Home

learning kinetic monte-carlo simulator charles l. thornton a

1. KINETIC MONTE_CARLO STEP16m ON Cu CU111 11 11 2003 new gen initial configuration file clust9i abk file of results clust92 r file with movie name or none clust92 m file for result configuration clust92f abk file with config database baza_c92 output sample for NEB and current config no update database yes analysis of 3d index in pattern recognition no file for output event_stack name or none stack_c92 output of config yes or no yes file with control parameters output clust92 con file with substrate config for NEB cuilir sub file with event statistics clust92 st file with energy distribution clust92 en output information about nonzero classes YES or NO yes number of time steps 1 e7 interval of information output in one run 1 1e4 number of static layers 2 temper K 500 interval for event_stack output lt 10000 9000 initial seed for random generator ISEED any integer 6543 Figure 1 5 The main configuration file step16 dt a global parameter in the SLKMC simulation The text in the first 60 columns of each line is not used by the software This allows users to enhance the readability of the input file The values starting at column 61 and extending to the carriage return are read in a specific order by the simulation software Notably this configuration file references clust9i abk as an initial configuration file baza_c92 as a database file and culllr sub as a substrate configu
2. At each step the simulation selects one process to apply to the model A process or transition represents a transformation that the sim ulator will apply to a particular configuration of atoms the energy threshold that the transformation would have to overcome and the change in time represented by the application of the process A configuration is a group of active atoms on the substrate When a configuration of atoms undergoes a process the positions of the atoms within that configuration are updated according to the rules defined in the process Figure 1 1 is an example of a configuration an a KMC simulation In the KMC model shown the central atom has a 36 atom neighborhood arranged in a hexagon The empty circles represent locations in the configuration that could contain atoms but do not the dark circles represent other atoms within the configuration The light gray circles in Figure 1 1 represent the underlying substrate Because the central atom lies above the substrate in such a way that the substrate forms a triangle pointing upward this atom is said to be lying on an face centered cubic FCC site as well as all of the other active atoms in this example More complex models can represent atoms lying on different kinds of sites within one configuration Figure 1 2 is an example of a process in a KMC simulation This process is one of APLELORLO LORO PAL HLH Led LOPD ode 09 gt 909907 lt 4 Figure 1 1 Exam
3. Rahman T S Kara A Trushin O Cluster Diffusion and Coalescence on Metal Surfaces Applications of a Self Learning Kinetic Monte Carlo Method MRS Proceedings 2004 Shim Y Amar J G Synchronous Relaxation Algorithm for Parallel Kinetic Monte Carlo Physical review B 71 115436 2005 Rahman Kara Karim Private Correspondence email rahman phys ksu edu 57 8 10 11 iii 12 at 13 14 15 16 Wallentine V Ramachandran A Private Correspondence email virgQcis ksu edu Gamma E et al Design Patterns Elements of Reusable Object Oriented Software Addison Wesley 1997 VTK Home Page The Visualization Toolkit Kitware Inc 14 Apr 2005 lt http www vtk org gt Kitware Inc Leaders in Visualization Technology Kitware Home Page Kit ware Inc 14 Apr 2005 lt http www kitware com gt Download Java 2 Platform Standard Edition 5 0 Sun Microsystems Inc 2005 lt http java sun com j2se 1 5 0 download jsp gt Wilson S and Kesselman J Java Platform Performance Strategies and Tac tics Sun Microsystems Inc 2001 ClassLoader Java 2 Platform SE 5 0 Sun Microsystems Inc 11 Aug 2004 lt http java sun com j2se 1 5 0 docs api java lang ClassLoader html gt Java Web Start Sun Microsystems Inc Oct 5 2004 lt http java sun com products javawebstart developers html
4. Click the button labeled Simulate in the left pane This will bring up the simulation execution view shown in Figure A 5 No simulation server is currently available so Figure A 5 The simulation execution view the only option is to load results from files Click on the button labeled Start to begin reading results data into memory The status field and progress bar will provide feedback about the status of data retrieval 63 A 5 Simulation Analysis As soon as the database is loaded the button labeled Results in the editor selector pane will become enabled Click the Results button The results view shown in Figure A 6 will be shown in the editor view From this view we can launch the 3D Figure A 6 The simulation results and analysis view results view or analyze the database statistics Click the button labeled 3D Results to bring up another dialog While the 3D subsystem prepares the scene a Rendering message is displayed however once the view is ready to use the message will be replaced by the simulation model Figure A 7 shows the progression of the 3D results view Once the view is ready we can Figure A 7 3D results view manipulate it using the mouse We can rotate the camera angle by left clicking on 64 the 3D window and dragging the mouse We can translate the camera position using the shift key along with a click and drag of the mouse Finally
5. Lear hcas Sik elon d ACN Data Model sas Hs Bei Pate te eg ae er BED EOS 4 1 2 Planned Scalability et a 4 Geen Je Bete Be Be BOE Pee 4 1 3 3D Subsystem AI ee o Ae ASMIE SMS id e o E reine rs E Ses 4 2 1 KMC Mesh Architecture Lar Deployment stranek Ree ake ee ge AY a la ae E Results Ok User Acceptances dnit gyi Bae A le oS deel le Sherk Ae RR be PE eo 5 2 Fut re Works 22224 a AS A oh is A eee ae amp 5 2 1 Enhancing KMC Vis A a Sgn oe ae eee 5 2 2 Enhancing KMC Mesh o 5 2 3 Extending the Core Simulation Code KMC Vis Walkthrough A a O E E E 111 38 38 39 42 43 45 45 47 50 ol 53 53 54 59 59 A 2 Getting Started a o ook sek Ce Sia ee Sia ag eB eg 61 A 3 Opening a Main Configuration File 62 AA Loading Simulation Results gt brad da ab Ane ale A Algal 63 Aor A oe aes te Gai ae AGAR Gude wwe 64 B KMC Mesh Walkthrough 67 B 1 SG eG hie Started ii ar A a AA ath Gs agen aoe ates 67 Bid Orestine a Mesh tiis srate a ES 68 Bid Savins eses ae es ea De aah te oe a Ee a ee 71 C GUI Evaluation by Users 72 List of Figures 1 1 1 2 1 3 1 4 1 5 1 6 1 7 1 8 1 9 2 1 2 2 2 3 3 1 3 2 3 3 3 4 Example of a KMC configuration oaoa 4 Example of a KMC process 2 oaoa aaa a Pa 5 Pseudocode for simulation initialization 6 Pseudocode for main simulation loop 7 The main
6. a coalescence problem used to verify the version of KMC underlying this research 1 2 SLKMC Technical Overview In order to understand the flow and data structures used by SLKMC it is first necessary to define a few terms Definition Let HashSet be a data structure with the following properties e Maintains a list of Key Value pairs e Provides get and put methods which access and modify the data structure e Does not allow duplicate keys e Each Value in the HashSet is a non empty Set of values Definition Let a Configuration be the arrangement of a central atom and its 36 neighboring sites Definition Let a Transition be the triple of initial configuration energy threshold and end configuration With those definitions it is possible to describe the behavior of the SLKMC soft ware Note that many other things are going on in the simulation but these algo rithms capture the core of the SLKMC implementation Figure 1 3 shows pseudocode for the algorithm that initializes the simulation In essence initialization simply pop Let tTable be an empty HashSet Let A be the set of all active atoms in the simulation for each atom a in A T getPossibleTransitions a for each transition t in T add the mapping t a to tTable Figure 1 3 Pseudocode for simulation initialization ulates the process list tTable with all of the possible processes that can occur Because tTable can map a single key or transition to
7. given as posf and DOS inal respectively Notice that no value is given for the duration of a process The same process can be applied over different periods of time Process descriptions are appended to the database any time the external function is required to generate them If a simulation is run using an initially empty database file a large part of the initial runtime is devoted to building up the database However two simulations with a common substrate configuration can share a database file In theory this would allow the databases to be linked to the substrate configurations in a technique that would eventually give a complete database However this technique has not yet been fully explored Unfortunately more complex models make exploring the configuration state space an increasingly difficult problem Initial Configuration File This file provides the simulator with the starting locations of each substrate atom and active atom A portion of an example initial configuration file is given in Figure 1 8 The first three lines in the initial configuration file allow users to add suitable documentation to the file They are not used by the simulation software Each following line represents a single atom in the simulation The first three fields contain floating point values that describe the position of the atom in R The next three fields describe a velocity vector for the atom In the initial configuration file this velocity
8. gt Java Web Start 1 5 0 Developer Guide Sun Microsystems Inc Aug 19 2004 lt http java sun com j2se 1 5 0 docs guide javaws developersguide contents html gt 58 Appendix A KMC Vis Walkthrough This walkthrough will demonstrate how to use the KMC Vis software This document is appropriate for both the Windows and Linux operating systems A l Overview Before we begin it is important to know exactly what role KMC Vis plays in the simulation A SLKMC simulation requires the following files the file names shown are those found in the sample data e Main Configuration File step16m dt A configuration file that specifies global parameters and the names of the other input files e Database baza_c92 The simulation database consisting of mappings between configurations and processes 99 e Substrate Configuration cu111r sub Configuration file for the substrate used in this simulation e Initial Mesh Configuration clust9i abk An initial mesh configuration Using these files the SLKMC software performs its simulation and dumps out a number of output files e Updated Database baza_c92 The updated simulation database e Database Statistics clust92 st Database process statistics for this simulation run e Event Stack stackc_92 A complete trace of all simulation processes This file is huge not currently used by KMC Vis and not included in the sample data e 3D Mov
9. in a tree configuration as shown in Figure 3 5 29 The roots of these trees are the atomic configurations from which processes may occur One level down is a complete list of transitions or processes that were initiated by the root atomic configuration Only those processes that have occurred during the simulation are listed The shown configurations are ordered by the frequency of their aggregate processes Because there is no natural labeling other than that applied by a human analyst the most meaningful label available for the tree is the aggregate frequency of a configuration or the frequency of a process One time consuming research task is identifying symmetrical processes KMC Vis implements a novel tool to assist in the discovery of such processes By selecting a process then clicking on the Find Symmetries button a user can launch the sym metry detection tool Figure 3 6 Within the symmetry detection tool the reference Find Symetries ref Threshold Energy 0 322 0 001 0000000000000009 C Symetrical Figure 3 6 Detecting process symmetries in KMC Vis process is indicated at the top of the process list and candidate processes are listed underneath Candidates are selected using the following qualifications 30 e The threshold energy AE must be within 0 01 of the reference process s thresh old energy e Only processes that occurred at least 1 of the time are considered Selecting a candid
10. it is possible to zoom in and out using the right mouse button and dragging forward zoom in or dragging backward zoom out Take a moment to experiment with the mouse controls The zoom and focal point can be reset to default by pressing the R key Play the simulation moving using the controls at the bottom of the 3D windows Playback can be paused at any time using the Pause button When a playback is paused it is possible to use the Step Forward and Step Back controls These allow you to advance or rewind the simulation one step at a time The Stop button causes the simulation playback to cease and the current frame to be reset to the first frame in the simulation movie Other controls are available on the right side of the window to control visualization quality and atom size Close the 3D results window and return to the top level results view Common database processes in terms of frequency are listed in a tree If you click on an entry in the tree you will not see a process rather you will see a configuration that can result in several processes If you expand the tree for that configuration and click on one of its children you will see arrows added to the previously clear view of the configuration These arrows show where atoms within the simulation are moving during that process Also the threshold energy for the process is shown in the Delta E field Now that a process has been selected the Find S
11. low medium and high settings for atom resolution The lowest available setting allows low end machines to render the 3D visualization Figure 3 8 Low medium and high resolution atoms with acceptable performance where as the highest setting is useful for presentation graphics The default resolution is medium It is possible to export presentation graphics of a particular 3D scene using the Save Image button This will open a file chooser where users can name and save their images Atom sizes can be adjusted via a drop down menu similar to the one for resolution 3 2 KMC Mesh KMC Mesh is a mesh generation program available as either a stand alone appli cation or from within the KMC Vis software It allows users to create and modify configurations of atoms in a visual interface Figure 3 9 is a screen shot of the ap 32 plication in action The following sections will discuss the features supported by the Figure 3 9 KMC Mesh screen shot version of KMC Mesh used in this work A walk through of the software is available in Appendix B 3 2 1 Mesh Creation A mesh in a KMC simulation is two complete substrate layers underneath a third layer that contains some number of active atoms The correct definition of a mesh requires the correct relative placement of each atom in every layer There are two ways to begin working with a mesh in KMC Mesh Figure 3 10 shows the File New dialog which allows t
12. of the software were solicited from users of SLKMC and KMC Vis The results were very encouraging Not only did the users believe that the software was a time saving tool but these reviews confirmed that users of the KMC Vis software were excited about the project and wanted to see it carried through to the next level Section 5 1 will provide details about the user feedback and how that validates this work Section 5 2 will discuss a variety of features that could be included in future versions of the software 50 5 1 User Acceptance Three researchers provided evaluations of this work The entire text of these letters appears in Appendix C of this document The first letter the text of which is in Figure C 1 was provided by Altaf Karim Altaf is a graduate research assistant at the Kansas State University KSU physics department His formal area of research is condensed matter physics and he is actively using the SLKMC software to examine the behavior of atomic islands on substrates The second letter of support Figure C 2 was provided by Abdelkader Kara Ph D Dr Kara is a research professor at KSU specializing in solid state physics He is also an active researcher in the SLKMC project In addition to traditional usage of the KMC Vis software he has used it during a conference presentation to enhance a talk about the SLKMC system The third evaluation Figure C 3 was written by Talat Rahman Dr Rahman is also a professor at KSU she
13. pairs While any user with a text editor can manipulate this file there are a number of problems with generating input in this way The first problem has to do with assumptions Without a knowledge of the underlying software system a user might assume that the input file was actually a sequence of key value pairs This is not the case The first sixty characters of each line provide a description of possible values to guide the user but these values are read by the simulator which assumes they are 17 in the proper order Even more disturbing is the realization that once the proper ordering of these lines is lost it could require a backup example or a knowledge of the underlying code to repair the error Another issue when directly interacting with the text file is validation Details like the valid width and format of floating point values entering boolean values and knowing what strings are valid input all require thorough documentation and expe rience to avoid potentially time consuming headaches Also simple typographical errors can occur in directly edited text files A final complaint against direct text editing is presentation Issues such as logical grouping and warning against underdeveloped options can only be addressed in the most rudimentary fashion within rigidly defined text input file However techniques are available to address all of these issues with a modern graphical user interface 2 2 2 Mesh Generation In the ma
14. the File New tech nique common in modern software Once loaded all the properties within the main configuration file can be edited from within the user interface Figure 3 1 shows the input options editor Every parameter in the main configuration file related to sim ulation input as opposed to those related to simulation results is editable within this view Categories such as Simulation Input Simulation Options and Database Options have been created to help guide the user The editor also validates all inputs prior to saving the file Figure 3 2 shows what happens if the user neglects to enter a mesh configuration file The software issues a specific error message and highlights the offending field To simplify the process of entering file names small folder icons next to fields requiring file names allow the user to open a file chooser and select a file This provides both convenience and protection against entering invalid file names 25 KMC Vis step16m dt Output Options Simulation Input Simulation Name Cu CU111 11 11 2003 new gen Initial Configuration clust9i abk gt Launch KMC Mesh oe Process Database baza_c92 E e NEB Substrate Config cu111r sub e Simulate J P Simulation Options Database Options Duration MC Steps 10000000 V Enable Dynamic DB Update Output Freq 11000 Enable 3D Pattern Recognition Stack Dump Freq 9000 Tem
15. time 1 2 1 Running Simulations This section will give an overview of each required file It is not intended to serve as a complete user s manual for the SLKMC software To run a SLKMC simulation the user must create four files A top level configuration file provides some global values and provides the names of other configuration and results files This configuration file references the other three required files These files are the e Process Database e Initial Mesh Configuration e Substrate Configuration These three files can be seen as representing the state of the simulation As the simulation runs checkpoint files are output at a preset interval The simulation can be rerun from any of these checkpoints or modified to create new simulations However since Kinetic Monte Carlo simulations are inherently non deterministic two simulations initialized from the same state are not guaranteed to follow the same path or reach the same final state Even though the main configuration file provides an initial random number generator seed there is not currently any way to capture the state of the internal random number generator at each checkpoint Main Configuration File This file sets a variety of global variables for the simulation and gives the names of other input and output files An example of the main configuration file is shown in Figure 1 5 The first line is a title for the simulation Each following line represents
16. until recently Now it is being solved to some extent by making use of the SLKMC User Interface It offers the user an essential tool to accumulate and categorize atomic processes on the basis of their energy barriers In short Charlie s work has expanded computational resources and computer graphics directed at users needs and users interaction with our SLKMC software Sincerely Altaf Karim Figure C 1 Review from Altaf Karim 73 Professor Virgil Wallentine Computing and Information Sciences Head Kansas State University Manhattan KS 66502 Dear Professor Wallentine This letter is to report on my experience using the software KMC GUI developed by Charlie Thornton as part of his masters thesis Let me say first that the final product is very much useful for the researchers involved with kinetic Monte Carlo KMC simulations In these simulations the researcher spends quite sometime in the pre and post simulation either preparing the system to be tackled or analyzing the various products spilled out by the simulation Note that in a typical KMC simulation the evolution of a system containing from a few to hundreds of thousands of particles is evaluated millions to hundreds of millions of times The KMC GUI consists actually of several components For the pre processing of the simulation the researcher can automatically generate a system and store it Note that before the building of this software every system
17. would be a small matter to generalize atom parameters and allow the user to customize how they are written out 54 The visualization window in KMC Mesh would benefit greatly from a few small enhancements The x y and z axes are not currently labeled on the 3D axis This can create confusion when rotating the view A related enhancement would be to include certain preset views for resetting the camera position Currently users can return to a 0 0 1 viewing angle using the Reset View toolbar option but it would also be nice to be able to reset the view to a 1 0 0 camera position 5 2 3 Extending the Core Simulation Code A detailed discussion of a distributed SLKMC simulation server is outside the scope of this document This section will discuss some of the features that would be desirable in such a software tool for optimal integration with the user interface software Ideally the computational node responsible for execution of the simulation would not be present on a user s workstation This allows the computationally heavy element to be executed on specialized hardware and prevent the user from undergoing the overhead of database files and results storage It should be possible to submit and track simulation jobs from within the KMC Vis software This requires that the server provide querying mechanisms for the current state of a user s simulation It should also be possible for the user to control the simulation database u
18. 8 is the Java Web Start file that launches the KMC Vis application from within a web browser This text is in a file named kmcgui jnlp and is used simply by publishing this file and the dependent jars on the web In order to package java class files and native libraries they must be placed in signed jars The process for creating web start ready jar files is described in the Java Web Start Developer Guide 16 Java Web Start enabled programs are launched using a java virtual machine within the user s web browser If that JVM is not in tact user s will need to fall back on 47 lt xml version 1 0 encoding UTF 8 gt lt jnlp spec 1 0 codebase http www cis ksu edu c1t3955 kmcgui href kmcgui jnlp gt lt information gt lt title gt Kinetic Monte Carlo GUI lt title gt lt vendor gt Charlie Thornton lt vendor gt lt homepage href index html gt lt description gt Kinetic Monte Carlo GUI lt description gt lt description kind short gt A visualization for kmc stuff lt description gt lt offline allowed gt lt information gt lt security gt lt all permissions gt lt security gt lt resources gt lt j2se version 1 5 gt lt jar href kmcvis jar gt lt jar href vtk jar gt lt resources gt lt resources os Windows gt lt nativelib href libs_win32 jar gt lt resources gt lt resources os Linux gt lt nativelib href libs_linux jar gt lt resources gt lt appli
19. A GRAPHICAL USER INTERFACE FOR THE SELF LEARNING KINETIC MONTE CARLO SIMULATOR by CHARLES L THORNTON B S Kansas State University 2003 A THESIS submitted in partial fulfillment of the requirements for the degree MASTER OF SCIENCE Department of Computing and Information Sciences College of Engineering KANSAS STATE UNIVERSITY Manhattan Kansas 2005 Approved by Major Professor Virgil Wallentine ABSTRACT Kinetic Monte Carlo KMC is a non deterministic computational technique for simulating atomistic motion These simulations can be used to predict the formation of substances at the atomic level This work describes a graphical user interface for a variant of the KMC simulation technique called Self Learning Kinetic Monte Carlo SLKMC which is currently under development We look at the SLKMC algorithm as well as the steps users take to extract useful information from the simulation data Then we look at potential ways to enhance accuracy and productivity during the model description and analysis phases of simulations The user interface described in this work includes support for the creation of initial condition data via mesh gener ation and global constant editing It also provides improved support for simulation results analysis Analysis features include animated 3D model visualization and statis tical data representation The architecture and implementation of software designed to carry out these enhancemen
20. ames for output files and select optional output options Validation similar to that found in the input options view protects the user from entering invalid output options Folder icons are not present to assist the user s file selection in this view because output files do not necessarily exist prior to execution of the simulation Once the user has edited the current configuration it can be saved using the File Save As technique common in modern software At this time the data will be validated and assuming the data survived validation the Simulate option will become available on the left editor selection panel 27 Notably absent from the pre processing feature list is creating three of the four required files from scratch This support has not been included in the version of the code discussed in this work Clearly this feature is important and must be added before the client software is ready to be used outside the KMC research group Section 5 2 discusses this and other future work for the project 3 1 2 Running the Simulation Once a valid main configuration file is in place the client software can retrieve simu lation data Figure 3 4 shows the Simulate editor Since this version of the software KMC Vis step16m dt Simulation Results Source Read Results from Local Files Simulation Progress Status loading database Poges C EE Figure 3 4 The simulation execution view p
21. anges are immediately applied to the mesh visualization shown in the left of Figure 3 9 This view provides a 3D representation of active mesh Atoms in the mesh are color coded according to state Active atoms are yellow inactive atoms are blue and disabled atoms are shown in a semi transparent white This side by side arrangement provides the user with a sense of perspective and orientation while working with a mesh Users have complete control over the camera position and zoom factor used in the 3D view via the mouse A Reset View button is available on the toolbar to return the camera and zoom to its default position Another feature of the 3D view is a highlighted box indicating the current editing layer As the user changes the editing layer using the Layer Up and Layer Down tools within the layer editor a translucent yellow box travels up and down the visu alization indicating the selected layer This option can be toggled on and off using a 36 toolbar button 37 Chapter 4 Software Architecture This chapter is a high level discussion of implementation details of the KMC Vis and KMC Mesh tools A section describing the Internet deployment scheme will follow the architectural discussion All of the code for this project was written using version 1 5 of the Java Development Environment The 3D data visualization libraries were written in C The interaction between the two was handled via the Java Native Inter
22. ate symmetry brings up a visualization of the candidates process in the tool s window Comparison information is displayed at the top of the window Because the reference process is still displayed in the results view of the main appli cation window it is possible to view both the candidate process and the reference process at the same time A checkbox can be clicked to specify that a process is sym metrical Once the user has checked all relevant processes and dismissed the dialog the database statistics tree rebuilds itself with a process s symmetries as child nodes This affects the overall frequency of the parent configuration and the root nodes are reordered Figure 3 7 shows an example of the 3D results view Viewing an animated movie MC Step 343 Total 908 Options Detail MED Size MED vw Save Image Figure 3 7 3D simulation history analysis of simulation progress allows researchers to identify windows in time where major changes occur as well as simply understand the progress of the simulation These 31 windows help to focus other types of analysis such as center of mass tracking Movie like buttons such as stop play and step are available to allow the user to easily move back and forth through the simulation movie Users have full control over the viewing angle and zoom within the 3D window It is also possible to change the resolution of the atoms in the movie Figure 3 8 shows the difference between the
23. blisher authenticity can not be verified The security certificate was issued by a company that is not trusted J The security certificate has not expired and is still valid More Details Details Figure A 2 Grant the application permission to run 61 A 3 Opening a Main Configuration File You will be greeted by the blank screen shown in Figure A 3 We can now load the Figure A 3 You will be greeted by a blank screen sample data we retrieved earlier Click the File Open menu item and browse to where you extracted the sample data directory Open the main configuration file step16m dt Once the main configuration file is loaded a variety of data is displayed in the tabbed panel on the right shown in Figure A 4 All of the data in the main config Figure A 4 The main configuration file editing view uration file can be edited in the Input Options and Output Options tabs of this 62 view Take a moment to flip back and forth between the views by clicking on the different tabs If you change any of the input or output data a modified main configuration file can be written using the File Save As menu option In this case we will be using the results from a previously run simulation so any changes will not affect the output however modifying any of the input or output filenames make it impossible for KMC Vis to load the results A 4 Loading Simulation Results
24. cation desc main class kmcvis Run gt lt jnlp gt Figure 4 8 The JNLP file for KMC Vis the zip files provided on the same web page There are a few options in kmcgui jnlp that are worth emphasizing It is a XML formatted text file The codebase option at the top tells the JVM where to find the jars listed in the resources section this value must be correct or they will not be found Within the security section the all permissions option is listed This option allows the application to interact with the user s hard disk at the expense of a launch time warning message In the field lt j2seversion 1 5 gt the plus sign 48 indicates that any higher version of the Java VM is acceptable however if the user has a lower version the listed version will be downloaded The resources section allows us to specify what jars are downloaded for specific operating systems Details for these options and everything else in the file can be found in the developer guide 16 49 Chapter 5 Results The most important factor to determine the success of this work was that the users of the KMC Vis software would accept and use the software This software was introduced to users that were already experienced in managing their data load They had software data analysis tools that already provided a means of computing any of the information accessible through KMC Vis In order to measure the success of KMC Vis reviews
25. class extending the Conf interface must be written This class must provide a mechanism to manipulate and retrieve the shells of atoms surrounding the central atom The second step is to create a version of the GridSite View class that understands how to work with the definitions of shells found in the new configuration Since there is no plan to continue supporting older versions of SLKMC since the hosted executable can be kept up to date there is currently no motivation to maintain older versions of atomic configurations Simulation execution support will need to be expanded to provide several different variants The current technique simply loads data files from the user s hard drive Additional support for a remote server and local execution would be useful features Figure 4 5 shows the architecture of the simulation run data the java class is RunSim 42 Data The data source for a simulation is an abstraction that can be implemented interface Data Source E gt Figure 4 5 Simulation execution data model in any number of ways The interface is called S mDataSource within the code and requires the following methods e getStatus String Return the current status of the simulation e start void Begin the simulation e stop void Cancel the current simulation Changes to the KMC Vis data model are performed internally using the producer consumer relationship described earlier This simple m
26. configuration file step16 dt 9 The databasefFile baza c92 es uti a RA 10 Database entry format sde sa ple teh whee tes tte fee ele 10 The initial configuration file stepl6i abk 12 Coalescence validation simulation results snapshots from Rahman et aa de tol e La rd a ee 13 Two symmetrical SLKMC processes eae 21 Stack file format AA A A A ee a 22 Stack file format key ups e Eo a Seale rd rs LES 22 Editing input options se cas io os a aa 26 Validation of input options 24200 al ae ed 26 Editing output options ooo a y dr a ara do a 27 The simulation execution view ooo a a a 28 3 9 3 6 3 7 3 8 3 9 3 10 4 1 4 2 4 3 4 4 4 5 4 6 4 7 4 8 A l A 2 A 3 A 4 A 5 A 6 A 7 A 8 B 1 The results view Be oS a AS aha Sak Soins ie hn NO Detecting process symmetries in KMC Vis 3D simulation history analysis och a 404 e ale aa A leas Low medium and high resolution atoms KMC Mesh screen shot o o e Creating a new mesh A ts gees ee ee ace ae eed The data model used by KMC Vis 004 Use of the publisher subscriber pattern in KMC Vis The publisher subscriber pattern Opening a configuration file The publisher subscriber pattern Updating a button s status Simulation execution data model The D pipelines ica o Ud le El a OP ees The KMC Mesh data mode
27. couple individual editors from the main application A common disadvantage of working with JNI libraries is the performance overhead Native method calls in Java are significantly slower than non native calls 13 VTK overcomes this problem by allowing the programmer to set up a scene once on the Java side then perform most of the computation within the native implementation This coarse grained architecture minimizes the JNI calls and provides tolerable 3D performance within Java Another advantage of working with VTK is the high level library support for 44 scientific data No code within the software created for this project uses primitives like triangle strips to render surfaces Rather it uses a data mapping and rendering pipeline defined by the VTK libraries that is optimized for scientific data Before attempting to modify the 3D code within KMC Vis or KMC Mesh developers would be well advised to examine the documentation at VTK s web site 10 because very few of the OpenGL variety of primitives that may seem familiar are present 4 2 KMC Mesh KMC Mesh is an extension of the KMC Vis software It consists of 16 additional source code files in two packages It also relies on code within other KMC Vis pack ages Because of the editor intensive nature of KMC Mesh it requires the finer grained data model observer system alluded to earlier The following section will discuss the data model used in KMC Mesh and also provides a rough g
28. ction the mesh editor view right hand side of Figure 3 9 34 provides several tools e Layer Up Changes the current editing layer to the next available layer in the zt direction This option is disabled when the top layer is selected e Layer Down Changes the current editing layer to the next available layer in the z direction This option is disabled when the bottom layer is selected e Reset View Resets the zoom and drag settings to the defaults e Zoom In Increases the zoom factor within the editor view e Zoom Out Decreases the zoom factor within the editor view e Drag Tool Default Allows the user to translate the editor view This is convenient when the current zoom obscures portions of the current layer e Make Layer Inactive Sets all cells in the current layer to the inactive state This is useful for creating substrate layers e Make Layer Empty Sets all cells in the current layer to be ignored 35 e Empty Atom Tool Allows the user to set individual atoms to the empty state e Inactive Atom Tool Allows the user to set individual atoms to the inactive state e Active Atom Tool Allows the user to set individual atoms to the active state These tools allow the user to quickly create custom simulation configurations without ever looking at the mesh configuration input file 3 2 3 Mesh Visualization As the user edits the mesh in the editor view those ch
29. d by a message or Event as listed in Figure 4 2 These events are essentially an enumeration within the kmevis model Event class By refraining from updating after irrelevant events subscribers can significantly reduce any performance penalty they may have incurred from hearing irrelevant messages Figure 4 3 shows an example of the producer consumer pattern within KMC Vis The code shown in Figure 4 3 opens a configuration file The configuration file is loaded and then inserted into the model Notice that the notifyObservers call is made at this level rather than within model itself If numerous modifications need to be made to 40 URL res jfc getSelectedFile toURL Config cfg Config load res Model getModel O setConfig cfg Model getModel notifyObservers Event CONFIG_OPEN Figure 4 3 The publisher subscriber pattern Opening a configuration file the model at the same time this technique prevents a serious performance penalty by waiting until all changes have been made then notifying listeners Unfortunately this approach weakens the abstraction somewhat by exposing the behavior to a modi fying class Figure 4 4 is an example of the consumer end of the communication The public void update Observable o Object arg if arg Event CONFIG_OPEN arg Event CONFIG_SAVE setEnabled Model getModel getConfigO isValidO else if arg Event CONFIG_CLOSE setEnabled
30. d versions of one another they represent the exact same underly ing physical process Accurate process frequency information can only be obtained by coalescing symmetrical processes into individual unique processes Analysis is carried out by building a list of all common processes then visually comparing them to search for symmetries Because of the extremely tedious and error prone nature of this work some sort of computational filtering prior to the need for human intervention would be very helpful Island Motion Analysis One highly useful result from an atomistic motion simulation is the mean square displacement MSD of the center of mass CM of an island of atoms 1 Running multiple simulations with slightly different initial conditions can lead to interesting results e g Raman et al 1 shows that clusters displace at dramatically different rates when they contain different numbers of atoms In SLKMC simulations this data is tracked via a binary stack file that has the format shown in Figure 2 2 21 A description of each of the fields in the stack file are given in Figure 2 3 In this At Raton n init posi po Miga pos init n pos final n Figure 2 2 Stack file format representation the byte length of floating point data is 8 bytes and the length of At Kaom n atom pos Elapsed time during this Monte Carlo step Central atom
31. e have created use the File Save As menu option A file chooser will allow you to specify the location of the file This file can be used as an initial mesh configuration file in SLKMC Any valid initial mesh configuration file can also be loaded using the File Open menu option 71 Appendix C GUI Evaluation by Users This appendix contains the full text of reviews solicited from the users of KMC Vis and KMC Mesh They support the claims found in Chapter 5 of this document 72 Dear Prof Wallentine I am writing this letter to express my views about the software Self Learning Kinetic Monte Carlo SLKMC User Interface developed by Mr Charlie Thornton This software is equipped with almost all essential tools required for the 3D visualization of atomistic processes on surfaces This nice and beautiful interface also enables a user to create customized crystal structures or atomic layers having ad atom islands and vacancies Prior to Charlies software I had been using another code that generates a crystal structure but it does not have the ability to create customized ad atom islands on atomic substrates Our SLKMC software generates a pile of large data files Analysis of these files is a tedious and time consuming task for a user This job of data analysis becomes more complicated when the user has to deal with files containing information like occurrence frequencies of atomic events Indeed this was a big problem
32. ection The default values will be sufficient for this example Click Ok A dialog will display the progress of the operation Now that the mesh has been generated we can experiment with the 3D view and 2D editor KMC Mesh allows you to edit a single layer at a time while visualizing the entire mesh A yellow box highlights the layer that is currently displayed in the editor The 3D view can be manipulated in the same was as the 3D view in KMC Vis Feel free to rotate the view so you can get a better look at the mesh If thinks get out of control you can always press the Reset View button on top of the 3D view Figure B 5 shows KMC Mesh after generating the mesh Notice that in this state no 00 RAK 808 00 Figure B 5 An empty mesh atoms are active or inactive they are all set to disabled or Empty Recall that 69 disabled atoms will not be written to the input file We are going to generate two substrate layers and then add some active atoms to the top layer To fill an entire layer with inactive substrate atoms click the Make Layer Inactive button in the 2D editor toolbar Is is a blue circle with an inset infinity symbol The result of clicking this button is shown in Figure B 6 Diz 00 44 88 000 Karer com Figure B 6 Filling the bottom layer with inactive atoms Now use the Layer Up 2D editor toolbar button denoted by an upward pointing arrow to shift t
33. entire configuration space and add each mapping to the process 15 database Pattern recognition schemes can be used to greatly reduce the space via configuration symmetries This reduction will become invaluable in the next gener ation of SLKMC which will use a 210 neighbor configuration model Some pattern detection technology currently exists in the software but a project is underway to add far more sophisticated artificial intelligence to the pattern recognition engine It is also important to mention that the symmetry detection must be perfect since even a small margin of error expanded over billions of Monte Carlo steps would seriously degrade simulation accuracy Assuming an effective system was in place no amount of symmetry detection could prevent occasional calls to the external function For this reason work is being done to parallelize the simulation Combinations of pooled and distributed technologies have been considered to off load the computation of potential processes 8 Because the computation of the process list for a particular configuration is both extremely expensive and completely independent of other computations it is a natural candidate for distribution How ever as previously discussed being forced to compute the external function is ideally a temporary problem The long term goal of any parallelization effort must be to im prove the performance of the individual time steps An interesting approach currently being exp
34. etPossibleTransitions function was invoked This function is the subject of scrutiny in the project because it is responsible for most of the up front computation and often much of the time consumption If a particular configuration has already been seen in the simulation getPossibleTransitions will simply return the set of transitions that can be applied to the configuration This behavior is the Self Learning aspect of SLKMC It is very powerful because there are only a finite number of configurations and even fewer reasonable configurations The database used to maintain information about past configuration to transition set mappings can also be reused between simulations so this presents an excellent oppor tunity to avoid the expense of the external function used to calculate the transition set for unknown configurations If a configuration shows up that has not previously been seen in the simulation it is necessary to invoke an external function to calculate the possible transition set This function uses techniques described in 1 and I will not discuss them in detail here The external function calculates all of the possible transitions that could arise from a particular configuration and it inherently takes a long time on the order of minutes to execute 7 The accuracy of the entire simulation is dependent on the resolution of this calculation Unfortunately this resolution is directly proportional to external function execution
35. face JNI package provided by the 3D library provider 4 1 KMC Vis KMC Vis consists of 66 Java source files collected into six packages Rather than document the detailed interactions between each of these files this section will dis cuss the data model used by the software and explain the communication paradigm Because the SLKMC software is a living project the 36 neighbor version became ob solete during this work there are particular portions of the code that were designed with scalability in mind This section will focus on those places in the software ar 38 chitecture Finally the 3D subsystem used in KMC Vis is the same as the one used in KMC Mesh A section is dedicated to explaining that system 4 1 1 Data Model The data model contains all of the data needed to write out the supported input files keep track of simulation progress and analyze the results Data internal to KMC Vis such as the preferred window size is also maintained in the data model but in the interest of clarity will not be described in this section Figure 4 1 shows an overview of the runtime data model used by the software The model is composed of Figure 4 1 The data model used by KMC Vis three components Config Run Data and Results Config maintains a record of all information related to the main configuration file for the simulation This includes the file names of the simulation input and output files The Run Data maintains informati
36. false Figure 4 4 The publisher subscriber pattern Updating a button s status code shown in Figure 4 4 is the update method for the action controlling the button that launches the Simulate view This button should only be enabled if a valid configuration file is loaded The code demonstrates how event filtering can be used to respond only to relevant events An understanding of the data model and the underlying communication mech anism should provide a programmer with the necessary understanding to read and modify the KMC Vis code Unfortunately this will be necessary for the user interface Al software to keep pace with the constantly evolving SLKMC system The next section will explain two of the places where we know the software will need to be enhanced lest it quickly become obsolete 4 1 2 Planned Scalability There are two obvious places to extend KMC Vis The first in the configuration model The 36 neighbor configuration scheme is already obsolete Current versions of SLKMC use a 210 atom neighborhood Also the data source for a simulation will certainly have to be expanded to include support for a simulation server Despite the emphasis placed on these two items the rest of the software was designed to be as convenient as possible to extend The only difference is the abstraction is already in place for these two components Implementing a new atomic configuration scheme requires two steps First a
37. for KMC simulations had to be built by hand and takes up much more of the researchers time The main advantage of having the KMC GUI software comes during the post processing phase The researcher may want to see a movie of the time evolution of the system during the whole or a part of the simulation Detailed analysis of correlation functions averages and distributions of a variety of physical properties have to be performed and the KMC GUI allows the research to access this information automatically at a significant gain in time and accuracy Finally 1 would like this software grow to satisfy other demands including extraction of information as the simulations go Sincerely Abdelkader Kara Research Assistant Professor Department of Physics Kansas State University Figure C 2 Review from Abdelkader Kara 74 Professor Virgil Wallentine Head Computing and Information Sciences Kansas State University Manhattan Kansas May 1 2005 Dear Professor Wallentine I am delighted to write this letter in support of the software KMC GUI that Charles Thornton has created as part of his Masters thesis project The software reflects a successful collaborative effort between members of two research groups one in CIS and the other in Physics We in Physics are excited at the final product because it enables us to create initial configurations for our simulations without the need to create a huge input file The ability to visual
38. from only a default set of configuration data stored within the software The two input files that can be generated via the user interface are the main configuration file and the initial mesh configuration Both of these files are fully editable from the user interface but the simulation database and the substrate configuration are not Support for managing both of these files could actually be provided through heavy re use of the existing code base a database manager would be similar to the statistics view and a substrate configuration tool would be almost identical to KMC Mesh Once these features were added users could reasonably expect to create the entire set of simulation input files KMC Vis supports one of three potential ways to run a SLKMC simulation load ing results from existing files It should also be possible to run the simulation on a SLKMC simulation server or simply execute a local SLKMC process and await the re sults Unfortunately this work was done prior to the existence of a simulation server and such support does not exist However even with that limitation it would still be possible to run simulations in a reduced performance mode using a local process A complete version of the software should support all three of these features unless the 93 local process is made impossible by the new architecture of the SLKMC executable The most fertile ground for improvement is in the simulation analysis stage As the sof
39. h of atomistic motion we are concerned with is the growth and movement of a substance on a substrate In order to produce materials according to certain specifications we must understand how individual atoms and atom clusters interact with one another on a substrate There are three approaches to modeling atomistic behavior Macroscopic modeling uses statistical descriptions of atomistic behavior to simulate atom cluster movement 1 Modeling can also be carried out at the atomic level by methods that simulate the movement of individual atoms There are also hybrid models that attempt to combine macroscopic and microscopic techniques 2 In this work we are only concerned with a model that simulates the interactions and processes of individual atoms Atomistic modeling has its roots in molecular dynamics MD simulations 3 MD simulations are based on Newton s laws and are totally deterministic The internal computations are straightforward but do not scale well to surface cluster modeling 1 Techniques have been developed that allow researchers to focus on interesting mo ments in the evolution of a system when clusters of atoms begin to merge or separate One of these relatively young methods is Kinetic Monte Carlo KMC 1 1 1 Kinetic Monte Carlo Kinetic Monte Carlo simulations use a non deterministic approach to model the move ment of individual atoms on a substrate The simulation proceeds in a series of Monte Carlo MC steps
40. he performance of the SLKMC simulation engine Research is underway to apply a parallel solution to this performance problem Since most users do not work on a multi processor computer it has been predicted that the final system will be a client server style architecture with the server commanding the power of many computational nodes The client software must then submit work and retrieve results from the server KMC Vis represents the client element of this architecture In recognition of the increasing popularity of the Linux operating system KMC Vis would ideally be available to users of both Windows and Linux based operating 24 systems This is easily accomplished using a cross platform language such as Java In order to smooth over interoperability issues in the future C may become a language of choice and in that case a multi platform toolkit for 3D visualization and windowing would be ideal The client software has been designed to make it as easy as possible to launch via the web Because mature versions those that interact with a simulation server will require an Internet connection this does not impose any penalties and enhances usability Java applications can take advantage of Java Web Start technology which allows the software to install and run with a single click Installers would have to be created for other forms of the client software 3 1 1 Pre Processing Main configuration files can be loaded into KMC Vis using
41. he user to create a custom mesh from scratch This method of mesh creation allows the user to specify the atom spacing and the number of atoms in the x y and z layers directions It is also possible to work with an existing simulation mesh Any valid initial configuration or final configuration of a SLKMC simulation can be loaded into KMC Mesh using the File Open command This makes it possible to create slightly different versions of the same simulation without directly editing a text file 33 Create Mesh x Mesh Size Atoms X 35 Atoms Y 35 Layers 3 Atom Spacing de 10 dy 1 0 dz 1 0 Figure 3 10 Creating a new mesh Any time a user would like to create a mesh suitable for a SLKMC simulation the File Save As option provides a simple way to write the mesh files The user is prompted for a file name and the current mesh is stored at that location Since the same format is used by SLKMC and the File Load option there is no need for a separate export function 3 2 2 Mesh Editing Any atom in the mesh can have only one of three states e Empty Disabled Disabled atoms will be ignored at file write time e Inactive Inactive atoms will be written at file write time with active flag value 0 e Active Active atoms will be written at file write time with active flag value 1 Editing a mesh is manipulating the activity state information for the atoms within a mesh To support this fun
42. ie clust92 m A file containing snapshots of the positions of the active atoms e Final Mesh clust92f abk A mesh configuration file suitable to be used as the input for another simulation KMC Vis does not perform any computation for the simulation rather it assists in the creation of input files and the analysis of output files 60 A 2 Getting Started We will need some sample data for KMC Vis A suitable zip file is available on the Internet at http www cis ksu edu clt3955 research php Download and extract the zip file at the link labeled sample data zip Remember where this data is we will need it later We can install and run the software from that same web page Click the link labeled kmcgui jnlp under the Run the Software heading This is a Java Web Start file and will run automatically Figure A 1 shows the application downloading necessary files at launch time Java Web Start _ Kinetic Monte Carlo GUI lt Charlie Thornton Loading libs_win32 jar from vwww cis ksu edu Yo Estimated time left 00 00 23 COT cancet_ Figure A 1 Running KMC Vis using Java Web Start Because KMC Vis requires access to your hard drive to read and write SLKMC files you will need to grant the application permission to run Do this by clicking yes in the dialog shown in Figure A 2 f Do you want to trust the signed application distributed by Charlie Thornton Pu
43. imulation is finished This approach is straightforward but it has some limitations related to parallelism and usability There are two unavoidable problems that will occur when running a KMC simu lation The first is a heavy CPU load and the second is the amount of output that will be generated The computational load is a product of the algorithms described in Chapter 1 The volume of output data is proportional to the number of MC steps requested in the simulation Small simulations produce hundreds of megabytes of out put large simulations produce gigabytes Ideally we want a way to execute problems quickly without losing the ability to analyze results as they become available 2 2 4 Simulation Results From the standpoint of results a KMC simulation is a stream of processes applied to particular atoms This information can be used in several ways e Visualizing the motion of atom clusters 19 e Analyzing the process statistics to explain the simulation e Numerical analysis of the motion of the atom clusters Meaningful data is collected using a combination of these three techniques The following sections will discuss the use of each of these techniques and the challenges that must be addressed for the KMC software to be more useful and accessible Visual Simulation History Despite the constant motion of a KMC simulation the key moments occur during a relatively small fragment of the simulation timeline 7 It is im
44. in configuration file each entry represented a conscious decision by the user This does not hold true for the mesh configuration file It is not realistic to expect a user to enter the locations of every atom in a 30x50x3 grid by hand 4500 lines of numerical text In determining the initial configuration for a simulation users should be able to provide a higher level description of a mesh then have a tool build it for them Tools such as this predated my research but there were a number of features that did not yet exist Users wanted the ability to populate meshes using pre existing island patterns 7 Ideally there would be a library of predefined islands that could be drawn from and added at particular locations or new shapes could be created Features that allow 18 creative control over the positions of individual atoms create a new layer of func tionality that wasn t present in existing tools They also require some form of visual feedback for the placement of the atoms 2 2 3 Problem Execution Once the correct input files are in place running the simulation is quite simple A compiled executable version of the KMC software will seek out the input files in the working directory read in the data and run the simulation Invalid input files will cause the simulation to prematurely halt As the simulation runs results data is continuously written to output files This data can be analyzed either immediately or after the s
45. ize the initial positions of the atoms and to change them at will in an instant is very handy and will be very helpful for our future work in the area It will also be very useful for external users of our code once it is ready to be placed on a dedicated web site But that is not all We are also very impressed by ability of the software to analyze the results of the simulations and to quantify the frequencies of various events together with visualizations of the processes events that participate in system dynamics In short 1 am impressed with the software that Charlie has produced and we expect to be using it for a long time to come It has certainly made our work much easier in terms of both input of data and understanding of the results We also expect the software to be very useful in our outreach activities and in broad dissemination of our results Sincerely Talat Rahman University Distinguished Professor Department of Physics Kansas State University Manhattan KS 66506 USA e mail rahman phys ksu edu Phone 1 785 532 1611 http www phys ksu edu rahman FAX 1 785 532 6806 Figure C 3 Review from Talat Rahmon 75
46. l o The JNLP file for KMC Vis 5 222548 it Ae Bode a Running KMC Vis using Java Web Start Grant the application permission to run You will be greeted by a blank screen The main configuration file editing view The simulation execution view o a The simulation results and analysis view 310 resulie View a A E A A A E Finding process symmetries o Running KMC Mesh using Java Web Start vi B 2 B 3 B 4 B 5 B 6 B 7 B 8 CA C 2 C 3 Grant the application permission to run 68 The initial KMC Mesh interface o o 68 The mesh creation dialog 4 Lu dd ale A ale a ad 69 An empty A ai es a a EG LOE GS BE 69 Filling the bottom layer with inactive atoms 70 A complete substrate st tas ee a es eas ee nage de ee 70 A complete mesh 50 5 4 is 40rd oe a See BS Be WAS Ewe ee 71 Review from Altaf Karim ae Pa ee aloe doe oN ee 73 Review from Abdelkader Kara ooa a 74 Review from Talat Rahmon Y asa A e we 75 vii Chapter 1 Introduction Physicists are increasingly using computational techniques to simulate the growth of substances under different conditions 1 Many of these techniques are complex and sharing the underlying software tools is not always enough to share the process itself However user interfaces allow u
47. lick on the kmcmesh jnlp link Java will automatically download the required libraries to run the application Figure B 1 shows the progress indicator for this action Because KMC Mesh requires access to your hard drive Java Web Start requires 67 Java Web Start amp _ Kinetic Monte Carlo GUI Mesh Generator Charlie Thornton Loading vtk jar from www cis ksu edu Read 1 6M of 5 7M 26 CCT Le Figure B 1 Running KMC Mesh using Java Web Start that you permit it to run Click yes in the dialog shown in Figure B 2 Warning Security Do you want to trust the signed application distributed by Charlie Thornton Publisher authenticity can not be verified A The security certificate was issued by a company that is not trusted 1 The security certificate has not expired and is still valid Figure B 2 Grant the application permission to run B 2 Creating a Mesh Once KMC Mesh is open you will be presented with the initially blank split view interface shown in Figure B 3 85 X24 80 000 Figure B 3 The initial KMC Mesh interface Click File New This will open the Create Mesh dialog shown in Figure B 4 68 This dialog allows you to specify the number of layers atoms in the x and y directions Create Mesh Mesh Size Atoms X Atoms Y Layers Atom Spacing Figure B 4 The mesh creation dialog and the atom spacing in each dir
48. llow the editing of individual element properties the model could easily be updated to include additional information In particular the velocity data in the substrate configuration file could be added in this way 46 KMC Mesh is designed to be independent of KMC Vis to allow stand alone exe cution This allows users to create view and modify mesh files without the overhead of running KMC Vis Of course the software is also available from within KMC Vis during main configuration file generation While the implementation of this feature is straightforward when Java is run normally it is unfortunately complex when the software is loaded from Java Web Start The code to correctly configure the class path and native library directories is in kmevis actions MeshGenTool This code should be reviewed when considering any changes to the launch paradigm 4 3 Deployment KMC Vis is available in two forms The first is a Java Web Start 15 enabled version of the software that can be launched from the web The second is a zip file containing all of the required jars and a batch file Versions for both Windows and Linux exist in both cases The benefit of the web start distribution is that the software will automatically update the user s version of the Java Runtime Environment 12 during launch unfortunately configuring such a system can be non trivial This section will describe the files used to provide this service in this research Figure 4
49. lored is to execute parallel updates localized to particular atom clusters 6 This technique promises to overcome the inherently serial nature of the single process per MC step simulation at the cost of very complicated division of labor 2 2 A Unified User Interface From a user s perspective there are three important tasks when running a simula tion The first is to describe the problem to the simulator This is accomplished by 16 generating the input files discussed in Chapter 1 The second step is to actually run the simulation and await the results As the simulation is running data is being gen erated Once the user has the required data the results must be analyzed in a useful way The following sections will discuss different elements of those three processes 2 2 1 Simulator Input A problem description consists of four input files e main configuration file e substrate configuration file e initial mesh configuration e process database substrate specific In general the substrate configuration file and the process database can be reused from existing sources Because the KMC project has not yet advanced to a point where users are creating custom substrates both of these files are assumed to predate the simulation The pre and post processing software is then focused on the main configuration file and the initial mesh configuration The main configuration file is a plain text file consisting of a number of key value
50. making use of the SLKMC User Interface It offers the user an essential tool to accumulate and categorize atomic processes on the basis of their energy barriers Dr Rahman also provided We are also very impressed by ability of the software to analyze the results of the simulations and to quantify the frequencies of various events together with visualizations of the processes events that participate in system dynamics Given the importance of the analysis function of the software this support was vital to validate the success of the project The non mesh data generation features and the simulation execution were not ad dressed in the reviews These features provided less functionality within the software than the mesh generation and data analysis tools More work needs to be done to determine the specific needs of users in these area as well as create an infrastructure to support software support of items such as remove simulation execution Details concerning future features are discussed in the following section 92 5 2 Future Work There is plenty of room for improvement in KMC Vis and KMC Mesh This section will discuss some of the features which ideally should be added to the project in order to more fully support SLKMC 5 2 1 Enhancing KMC Vis Currently it is necessary to use the results from a previous simulation to create input files using KMC Vis Ideally a user should be able to create an entire simulation model
51. many values in this case atoms the important data are the processes not the central atoms It is done in this way because we calculate the probability of a process independent of how many places it can occur We will discuss the get PossibleTransitions function later Once the simulation has been properly initialized the main loop will repeat until the desired number of MC steps have been calculated Figure 1 4 shows pseudocode for the main simulation loop At each MC step we select a weighted random tran Let transition chosenTrans getWeightedRandomTransition tTable Let atomList tTable get chosenTrans Let centralAtom be a random atom from atomList perform chosenTrans on centralAtom s configuration for each active atom a in centralAtom s configuration remove all mappings to a from tTable ts getPossibleTransitions a for each transition t in ts add t a to tTable Figure 1 4 Pseudocode for main simulation loop sition from our table of possible transitions The weighting is determined by the energy threshold of that transition Once we have a transition it is possible to ex tract the list of atoms whose configurations are valid candidates for that transition from tTable A random atom and its respective configuration is selected from the list of candidate atoms That configuration is transformed and the transition table is updated to reflect the new configuration In both pseudocode examples the g
52. o the next layer in the z direction Notice that the yellow box in the 3D view is now in the center layer of the mesh Set all of the atoms in this layer to inactive using the Make Layer Inactive button and shift to the top layer The current state of your mesh should be similar to that shown in Figure B 7 z Eje 252246 98 000 E Figure B 7 A complete substrate Now it is time to add some active atoms to the simulation Select the Make Atom Active tool denoted by a circle around a yellow dot from the toolbar This tool 70 will allow you to paint active atoms onto the layer one atom at a time Experiment with this by adding some atoms to the current layer You may notice that active atoms are difficult to see underneath the highlighting rectangle in the 3D view You can disable the layer highlight using the 3D view toolbar button denoted by blue and yellow layers Figure B 8 shows an example of the finished mesh 20 00a 0806 000 Figure B 8 A complete mesh The 2D editor toolbar also provides tools to zoom in zoom out and translate the view The translation tool requires you select it then drag the layer representation into the desired position The zoom buttons simply increment or decrement the zoom each time they are clicked There are also toolbar buttons to do variants of the tasks we performed during this walkthough B 3 Saving Mesh Data To write out the mesh w
53. odel allows great freedom in the implementation of a data source and exists to enable user interaction and feedback rather than to provide any hidden behaviors 4 1 3 3D Subsystem The 3D subsystem used in KMC Vis is an open source scientific data visualization package called the Visualization Toolkit 10 VTK released by Kitware 11 Because 43 of platform independence restrictions the Java language is poorly suited to directly implement any 3D toolkit VTK provides a Java Native Interface JNI layer to communicate with C code that in turn drives the user s high performance video hardware This relationship is illustrated in Figure 4 6 The Java code that connects Figure 4 6 The 3D pipeline to the native VTK libraries is generated automatically by VTK s build This code is stored in vtk jar and included on the project class path A strength of KMC Vis is lazy loading of the native libraries referenced by the code in this jar Because of how the Java Virtual Machine JVM loads classes into memory 14 the static loadLibrary calls in the VTK code would have two undesirable effects First the application would start slowly because the libraries would be loaded immediately Worse if there was a problem with the native libraries the application would not start at all KMC Vis is designed to use reflection to load all editors as they are requested to avoid the problems associated with loading the native libraries as well as de
54. on about the current simulation status Notably this component contains the data source for the simulation that will be used to generate the results information The results section of the model contains the mesh and movie data Strangely absent from the results data is the simulation database The database is actually stored at the top level within the model Statistics are applied directly to the database when the results are read in 39 Because of the need for loosely coupled communication within KMC Vis a publisher subscriber 9 design pattern was used to notify dependent components of changes to the model Within the Java language a framework for this pattern already exists in the Observable class and the Observer interface Figure 4 2 shows how this pattern was used within KMC Vis Any class extending the Observable class can be listened to Figure 4 2 Use of the publisher subscriber pattern in KMC Vis by a class implementing the Observer interface In KMC Vis only the top level of the data model was made observable The alternative was to make every element of the data model observable and then each editor and view component could subscribe only for relevant updates Such a fine grained solution would be very desirable if the data model were frequently changing but this performance benefit never outweighed the simplicity of a single publisher KMC Mesh does implement a fine grained solution Any change to the model is accompanie
55. perature K 5000 Inactive Layer Count 2 Random Seed 6543 Figure 3 1 Editing input options KMC Vis step16m dt Input Options Output Options Simulation Input Simulation Name Cu CU111 11 11 2003 new gen Initial Configuration ms gt LaunchkKMCMesh Stack Dump Freq Temperature K Inactive Layer Count Random Seed Figure 3 2 Validation of input options 26 A button labeled Run KMC Mesh can be seen among the input options KMC Mesh is a program specifically designed to assist in the creation of initial mesh con figuration files those ending with abk KMC Mesh can be executed from within KMC Vis or as a stand alone application Section 3 2 will cover this tool in more detail Similarly all parameters related to a KMC simulation s output can be edited via the Output Options view as shown in Figure 3 3 This view allows the user to KMC Vis step16m dt Input Options Output Options Results Files Config Simulation Trace clust92 1 Final Configuration clust92f abk oF ont eters clust92 con Simulate clust92 st Energy Dist clust92 en d Optional Files Movie File clust92 m Event Stack stack_c92 Non zero Class Info NEB and Config Samples Echo Configuration Figure 3 3 Editing output options choose file n
56. ple of a KMC configuration potentially numerous transitions the configuration in Figure 1 1 could undergo The arrow in this figure represents the movement the atom at its root will undergo In this particular example the central atom will move along the edge of the other atoms It is not uncommon for many atoms to move in a particular process At every MC step a process is selected in a weighted random fashion based on the energy thresholds of each potential process in the simulation This process will transform a small portion of the model and then the next process will be selected As there is no notion of convergence the simulation will run until a desired number of MC steps have been calculated As the atoms shift their positions the cluster shape its energy and dynamics change Tracking these changes is one of several ways to analyze atomistic simulations Using this information scientists can learn how different materials form and how to create desired behaviors The constant motion of these entities prevents the stagnation found in the older Figure 1 2 Example of a KMC process more rigid atomistic models Unfortunately this comes at the price of higher compu tational complexity and the need to specify more accurately the details of the system Efforts are continually underway to refine existing KMC approaches by increasing the resolution of the computational model The validation section later in this chapter discusses
57. portant to be able to monitor the simulation in a way that allows researchers to view progressively finer grained data and identify these important moments The mechanism used for this purpose prior to this research was a movie file that contains the positions of the active atoms within the simulation The position of each active atom was written to the movie file after a number of MC steps specified in the main configuration file Unfortunately this system fixed the granularity of movie output for an entire simulation Unless the user guesses correctly the first time finer granularity could only be obtained via additional executions of the software Process Statistics Analysis There are a variety of different types of database processes Rahman et al 5 contains a listing of 27 unique process variants discovered in a single experiment Identifying what kind of processes the system underwent and their frequency acts as an explana tion for the changes that occur within the simulation model during the key moments described above 20 A major part of analyzing this statistical data is identifying symmetrical processes 7 An example of two symmetrical processes is given in Figure 2 1 Since these two vvvy vyvvy OO YYYY YY bb Gib dab dah dab O O A Q 00 00000 O 0000 OOO 0000 O O Q Q Q Q Q Q on O Y O O o O Q O Y Figure 2 1 Two symmetrical SLKMC processes processes are rotate
58. ration and analysis tools from within the KMC Vis interface e KMC Vis will be the client in a client server architecture A high performance simulation server will handle computation e To accommodate the scientific community the software must run on both Linux and Windows operating systems 23 e Retrieval and execution of the client application via the Internet must be as simple as possible The remainder of this section will discuss each of these features and how KMC Vis addresses them Following sections provide details on the different components of the user interface A tutorial for KMC Vis is available in Appendix A Prior to this work their were a variety of FORTRAN programs written by the designers of SLKMC These programs allow their authors to automate some of the pre and post processing tasks involved in a simulation Unfortunately while this system brings enormous power to a small group of developers it creates a very sharp learning curve for new users KMC Vis is intended to be a single application that brings together all of the tools that the simulation software users need in such a way that their use is intuitive The implementation of KMC Vis in this research represents a first pass at this goal and was intended to demonstrate the feasibility of such a project Due to time constraints on this phase of the project it is not possible for the user interface to provide complete functionality A pressing issue is t
59. ration file These are the three required additional input files and we will briefly discuss each one Database File This file maintains the running record of all calculated database processes A portion of an example database file is shown in Figure 1 6 The first four lines allow the user baza for 7 atom at 500K cluster 7 1 2004 35 32 37 59 14 2051 o 1 0 670 1 1 4 3073 196608 1 0 052 1 1 2 2051 o 1 0 312 1 1 3 22 0 0 32 o 1 0 322 4 1 4 2 3 311 T FL 40 128 1 0 647 1 1 2 Figure 1 6 The databasefFile baza_c92 to insert comments describing the database and the simulation it was used in The entry format is shown in Figure 1 7 shell shellg shello n AE mi POS mig POS inal Figure 1 7 Database entry format The first three entries shell 4 shellg shellc in each line of the database refer to a particular configuration of an atom cluster Binary encoded integers are used for this representation to conserve space The next entry n gives the number of 10 different processes that can occur from that particular configuration Each process is fully described by the following entries The entry AE is the threshold energy for the i process The entry m is how many atoms within the shell configuration moved in the process The value of m will always be non zero For each moving atom we have an initial location and a final location These are
60. re dates any simulation server only the option to load results data from local files in available Options designed to assist the user in connecting to a simulation server are left disabled Clicking the button labeled Start shown in Figure 3 4 causes the software to attempt to load data from the database and results files listed in the main config uration files Attempting to load data that does not exist will result in an error 28 Progress messages are displayed to keep the user informed during what may be a lengthy loading process Once the simulation database is loaded and the other results information is avail able the Results option will become available in the left editor selection panel It is possible to view results while only a fragment of the simulation data is available to allow users to track the progress of simulations and view data as it emerges 3 1 3 Simulation Analysis Progressively more results data becomes available as a KMC simulation runs KMC Vis provides two major modes of analyzing this data Figure 3 5 shows the top level results view Database process analysis is available on the top level screen 3D KMC Vis step16m dt Simulation Movie Frame Count Frame Interval 3D Results Analysis Tools Create Data Table Figure 3 5 The results view simulation history visualization is also available in a sub window Database processes are presented
61. s position in k space Number of moving atoms int Initial position of the i moving atom New position of the i moving atom Figure 2 3 Stack file format key integer data is 4 bytes Entries in this stack file are written out every time step and custom data reader software has been created to perform analysis on the data After growing for millions of MC steps this simulation stack file becomes quite large Custom readers for the file format have been written by the developers of SEKMC but they do not allow users to target particular interesting portions of the data Ideally users should be able to plot the MSD for an island at the particular time window that is significant for the simulation without resorting to trial and error Also it would be very helpful if users were allowed to perform different calculations on this step wise data to generate a variety of graphs 22 Chapter 3 Software Description 3 1 KMC Vis This chapter will discuss the features in an application designed to address many of the productivity and usability issues raised in previous chapters The software is called KMC Vis because it represents a visual interface to the KMC package The objective of KMC Vis is to allow a user to prepare execute and analyze an atomistic simulation all within a single program Following that assumption there are a several important guiding characteristics e It must be possible to run all data prepa
62. s to control the level of detail presented to the users to provide graphical feedback to assist user input and to optionally automate complex analysis jobs all without reducing the power of the underlying software This research explores a pre and post processor for a atomistic motion simulator written in the FORTRAN programming language The user interface was written in Java and is completely separate from the computational software Communication is accomplished via the same text files the FORTRAN software uses to read input and produce output This means the physicists could continue work on their simula tion code without having to learn a new programming language or worry about any graphical techniques Development and adoption of the user interface was completely successful In the following chapters we will discuss the specification features and architecture of this work Verification of the success of this project has come in the form of reviews by users in the KSU physics department and their experiences with the software during normal use and during a conference presentation The full text of these reviews can be found in Appendix C The particular technique used by the underlying software for modeling atomistic motion is called Self Learning Kinetic Monte Carlo SLKMC This chapter discusses the atomistic modeling concepts and some of the details of this particular implemen tation of SLKMC 1 1 KMC Background The branc
63. sed in a given run This would allow both open Self Learning and closed databases to be used Also it should be possible to work with and contribute to an ever growing database of processes to improve future performance Pattern recognition plays an important role both in simulation performance and in database statistics analysis These techniques improve the performance of the sim ulation engine by preventing unnecessary calls to the external function The current 59 SLKMC implementation applies some rudimentary pattern recognition techniques but does not identify all symmetrical processes 7 As a result symmetrical processes leak into the database statistics necessitating the computer assisted symmetry detec tion tool discussed in Section 3 1 3 A computational layer that could quickly and accurately detect symmetrical processes would both improve simulation performance and completely eliminate a very time consuming analysis job 56 Bibliography 1 7 Rahman T et al Atomistic Studies of Thin Film Growth Proc SPIE Annual Meeting 5509 1 2004 Ratsch C and Venables J A Nucleation Theory and the Early Stages of Thin Film Growth Journal of Vacuum Science and Technology A Vacuum Surfaces and Films 21 96 2003 A Rahman Correlations in the motion of atoms in liquid argon Physical Review 136 A405 1964 Wolfram S A New Kind of Science Champaign IL Wolfram Media 2002
64. specializes in atomistic modeling and high performance supercomputing In their reviews all three users spoke highly of the KMC Mesh tool for generating initial mesh configuration files In support of KMC Mesh s visualization and editing 6 tools Altaf wrote I had been using another code that generates a crystal structure but it does not have the ability to create customized ad atom islands on atomic 6 substrates Similarly Dr Kara wrote before the building of this software every system for KMC simulations had to be built by hand and takes up much more of the researchers time Additionally Dr Rahman wrote The ability to visualize the initial positions of the atoms and to change them at will in an instant is very handy and will be very helpful for our future work in the area Section 5 2 2 discusses a variety of features that could be added to the KMC Mesh software l The analysis phase of KMC Vis also did quite well in the reviewing process In support of the analysis options offered by the software Altaf wrote the following Our SLKMC software generates a pile of large data files Analysis of these files is a tedious and time consuming task for a user This job of data analysis becomes more complicated when the user has to deal with files containing information like occurrence frequencies of atomic events Indeed this was a big problem until recently Now it is being solved to some extent by
65. the code into the hands of computer scientists for greater convenience in enhancing performance and usability would not be ideal For this reason the coupling of any software modules that address either of the second two areas must be very loose The remaining sections in this chapter will discuss areas of the SLKMC software that need to be improved This thesis focuses on the effect of enhancing the software usability thus usability improvements are given considerable emphasis 2 1 Simulation Performance The development of more accurate simulation technology is entirely dependent on the performance of the system Currently the accuracy of SLKMC is fully scaleable but growth in this area is stunted because simulations can take weeks to run From the main simulation loop described in Chapter 1 we can isolate the two performance bottlenecks in the software The first in computation of the external function Calculating the possible processes that a particular configuration can undergo takes a matter of minutes with the present computing technology The good news is this computation never needs to be repeated for a particular configuration This is the Self Learning nature of SLKMC as the database grows the external function is needed less often One technique that is currently being applied to this problem is pattern recognition 1 In the 36 neighbor configuration model demonstrated in this work it is possible to simply explore the
66. this validation are exciting because 12 Figure 1 9 Coalescence validation simulation results snapshots from Rahman et al 1 they agree with previously collected experimental evidence 1 A detailed discussion of the validation simulation can be found in Rahman et al 5 13 Chapter 2 Problem Description The existing Self Learning Kinetic Monte Carlo SLKMC software is evolving in three key areas e Simulation Accuracy e Simulation Performance e Usability Efforts to increase simulation accuracy by increasing model resolution are continu ally underway In order to provide suitable performance for finer grained simulations optimizations are needed both in results data size and simulation running time Cur rent simulations produce 1 3 gigabytes of data and take hours to run This time can extend to weeks if the process database must be discovered by the system Usability is important if we intend to share not only the simulation results but the process itself Other scientists will find it much easier to pick up a simulation package if they do not need to concern themselves with the details of file formats or are forced to recompile the software to make minor changes 14 A key idea in this project is that physicists should continue to develop and main tain the core computational engine within the SLKMC software Their backgrounds give them tremendous insight into the underlying problem and shifting ownership of
67. ts is also discussed We assess the usefulness of the implementation of the software using reviews conducted by developers and users of the SLKMC simulator These reviews verify that the unified interface contributes to both the usefulness of the underlying simulation code and user productivity Contents List of Figures iv 1 Introduction 1 1 1 KMC Background dao le a da an e de O a 2 1 1 1 Kinetic Monte Carlo sio ai a aa 3 1 2 SLKMC Technical Overview to fee A A he 5 1 2 1 Running Simulations 2d 464 a a ee EAE e 8 12 2 Validation teas o Ad A A e 12 2 Problem Description 14 2 1 Simulation Performance e 15 2 2 A Unified User Interface RAE AAA ERO e 16 22 Simulator ap cit da ade ad eh a bes and 17 2 2 2 Mesh Ge nera tion ad e des amp e S a A 18 2 2 3 P oblem Executi a 4 a aneia ere ets ee e e p nd e 19 2 2 4 Simulation Results ss 424 44 5 25 a dt pate be ah 19 3 Software Description 23 ol E a enh fa oe core Bo og cee Bo ntl CaS Ae once i tS ak te SEN Os 23 ii 3 1 1 e ro 0 aa Ce ea ok en Sa ne ON 3 1 2 Running the Simulation a elie an ale pe eS 3 1 3 Simulation Analysis aaa a Al 32 E s sok aan ane ee ae ih he ec hs ane ENEA ae ees 3 2 1 Mesh Creation at be A de ee lS eke a A de 3 2 2 Mesh A A eas ania as te Gas Geni aoa aa dened 3 2 3 Mesh Visualization dto ee lees ROS Bee BOG EE ROUT Software Architecture AUN E uy Oe BS Beg e Se Beth Se Ge nth hg Be Rearlc
68. tware evolves both the users and developers find better understanding for how they can use it more effectively One key feature that is missing from the analysis stage is 2D graphing of equations such as MSD of the center of mass Another is results filtering to allow only a particular window of the simulation to be available for analysis Also the computer assisted process symmetry detection tool would benefit from a variety of features including process naming loading and saving processes to a file and statistical sorting tools for the database with and without symmetries Analysis features such as these are what makes the software truly useful and require the most attention 5 2 2 Enhancing KMC Mesh One of the strengths of KMC Mesh is that it can be generalized to many different mesh types rather than just 111 and many different formats It should be possible to modify the mesh type at any time during editing It should also be possible to plug in different file formats to allow the software to work with other applications that use a very similar MD style input file Another potential feature is property management for atoms in the mesh Cur rently an atom can have one of three states Extending KMC Mesh to also support substrate configuration files would be a matter of keeping track of velocity vectors in the x y and z directions Since this functionality would require more user interac tion than the mouse clicks currently required it
69. uide to a simple extension 4 2 1 KMC Mesh Architecture The KMC Mesh software consists of a data model peers for each object in the data model and a set of actions that modify the data An overview of the data and visual ization system is shown in Figure 4 7 KMC Mesh uses three data model components e Model The model contains at least one commonly three Layer object In addition to maintaining the list of layers spacing information and transient view informa tion e g the currently selected layer is also in the model 45 Figure 4 7 The KMC Mesh data model e Layer The layer structure contains an array of Element objects that represent all of the atoms in that layer It also maintains what type of layer to control x and y offsets e Element The element structure represents atoms in the simulation Each atom has a position on the x y plane and a type The type can hold one of three values empty inactive or active These values correspond to the type of atom Each of these components has a visual peer object within the VTK scene hierarchy These peers listen to their counterparts in the data model via the producer consumer relationship indicated in Figure 4 7 Any time a data element changes the visual peer is updated to reflect the changes to the user Currently the only data contained within the element structures is the state of the atom If the editor were updated to support right click context menus to a
70. vector is always zero The final field is a flag to indicate that an atom is 11 INITIAL CONFIGURATION 3214 14 102 2476 88 54905 8 348480 0 1600 3 0 00000 0 00000 0 00000 O O 0 0 1 27810 2 21373 0 00000 O O 0 0 2 55619 0 00000 0 00000 O O 0 0 Figure 1 8 The initial configuration file step16i abk either active or inactive Active atoms use a value of 1 for the flag They will be moved around by simulation processes Currently active atoms only reside on the top layer above the substrate atoms Inactive atoms use a value of 0 They represent the substrate and occupy the lower two layers of the simulation in a three layer model Substrate Configuration File This file provides the previously mentioned external function with information about the current substrate Its format is identical to the initial configuration file The only noteworthy difference between this file and the initial configuration file is that the velocity vectors are frequently non zero in the substrate configuration file because of the molecular dynamics equations used in that computation 1 2 2 Validation The SLKMC software has been validated against experimental data in a coalescence simulation In this simulation a large island of atoms is placed near a smaller island of atoms In time the smaller island will merge with the larger one 5 Figure 1 9 is a visual history of this simulation The results of
71. ymmetries button is enabled Clicking on this button will produce the dialog shown in Figure A 8 Select one of the candidate symmetries for the current process and click the Symmetrical check box For now don t worry if the processes are actually symmetrical Click the Ok button to exit the dialog The database process list will update to reflect the new 65 Find Symetries ref Threshold Energy 0 323 0 0 e fz Figure A 8 Finding process symmetries information If you open up the same process you will see that it now has a child process This was the symmetry we just added The symmetry can be removed by selecting the parent process opening the Find Symmetries dialog and removing the check from the candidate process 66 Appendix B KMC Mesh Walkthrough This walk through will demonstrate how to use the KMC Mesh software KMC Mesh is a graphical initial mesh configuration generation utility for KMC Vis The files generated by KMC Mesh are directly useable in SLKMC simulations This document is appropriate for both the Windows and Linux operating systems B 1 Getting Started We can install and run KMC Mesh with a single click Open a web browser and go to the web page at http www cis ksu edu clt3955 research php Under the Run the Software heading there is a link labeled kmcmesh jnlp This link points to a Java Web Start file that will trigger the application launch C

learning kinetic monte-carlo simulator charles l. thornton a

Contents

Download Pdf Manuals

Related Search

Related Contents