User's Guide VISIPART - Applicaties Helpdesk Water
Contents
1. Tfimona User s Guide VISIPART SIMONA report number 2008 01 Rijkswaterstaat Ministerie van Infrastructuur en Milieu User s Guide VISIPART Version number Version 1 3 June 2011 Maintenance t see www helpdeskwater nl waqua Copyright d Rijkswaterstaat Log sheet document date Changes with respect to the previous version version 1 0 16 01 2008 c77580 update of old documentation Visipart 1 1 19 10 2010 c3320 Support in Visipart for DDHOR simulation 1 2 21 10 2010 c3320 Converted to IZTEX 1 3 10 06 2011 c3320 Note on compiled version and removed empty ap pendix Version 1 3 June 2011 User s Guide VISIPART Contents 1 2 11 Directions for use of this manual Description of the menu items of Visipart 2 1 Starting and stopping Visipart ks R A X ACA X a E AECA A 2 2 1 File Import WAQUA enclosure o llle 2 2 2 File Apply partitioning a an A Bee Bee B a S BANE 2 2 5 File Load partitioning a k k k k k k e powder Auda bbe oan pee S 2 23 File Save partitioning exce ee Rie RARE RR URGE E at 22 0 Filej PODA xd xx ki ss k mex k k k dex 3 93 kak ad 22 4 EHE BRE ix Be B LR Beleen BA B Beeken B at S ae Boas v Se hen ain tna a ent iad a Sonne Sena Sle 234 Edit Undo Leik ede eee bee 3x4 bende oH 444 444 M T 2 34 Edit AE O ei ae dnt al PEET T n 2 4 1 View Zoom TB osa ga gogo egt2. HOW e hs HE EUR ACC OLOR de ZAZ View Zoom OUL a ks moe CA re Bae OK ROO WO o ROS x ek desc gui de dom dox ee ee ee X ac acer x 244 View ZoomFull 22592R RR RR E ss b oe MEER SANE no RE ES 246 View Show Grid ne as goo m Eh ROC OMA OCA ED ew OO OQ QV QV QV QN QN UA tA a E E HL E HP HW WW NY NY WY MN CONTENTS 2 4 7 View Show Curvilinear 5454 54 EO RE ROREM E REG EE 2 9 Options of the Options menu us Ros mox Ses 2 ee Boek Se Seow wees 2 5 1 Options Show estimated workload 2 5 2 Options Show Fullbox 2 5 3 Options Show Fill ratio eeen T ee a a ad eri crios men 0 OO 0 0 A A A A A A 3 Optimizing grid partitionings 10 4 Example for assigning processes in DDHOR simulation using Visipart 12 Version 1 3 June 2011 lii Chapter 1 Directions for use of this manual Chapter 1 Directions for use of this manual Visipart is a Matlab based graphical tool to assist in creating or modifying grid partitionings or grid decompositions for use with parallel computing or domain decomposition This document describes the use of the tool Chapter 2 describes the functions of the program and Chapter 3 describes a way to use the tool in order to optimize partitionings for a given simulation Version 1 3 June 2011 1 User s Guide VISIPART Chapter 2 Description of the menu items of Visipart 2 1 Starting and stopping Visipart Visipart can be started using the run proc
3. document about parallel computing and domain decomposition ugcou ple pdf For Visipart this configuration file is parsed by Matlab All keywords should be specified on separate lines in order for Matlab to be able to do this In this example we assume we have an initial partitioning for our DDHOR simulation containing three domains links rebo and reon So we need at least 3 processors to run this simulation Each domain has a different number of layers These 3 domains are visualised in Figure 4 1 To be able to assign processors to the domains we need the run times of each domain Note that when editing in one of the domains Visipart does not know about the other domains Normally a guess can be made from the number of grid points in each sub domain but in case of DDHOR it is not possible to determine the number of grid points in subdomains outside your own DDHOR domain To read these run times and show them we use Options Show Observed Run Times and select the wagpro m message file We end up with Figure 4 2 The Accept speeds button can now be used in the same way as for a parallel simulation except for the fact that Visipart does not know which domain is processed currently So if the Accept speeds button is used and we have a DDHOR simulation Visipart will ask which domain is processed see Figure 4 3 In this example we are editing the first domain and select the name links 12 Chapter 4 Exa
4. domain decomposition if the subdomain interfaces are far enough from each other and from the open boundaries When the option Options Check Domain Decomposi tion is selected a dialog box will pop up that reports potential errors an arrow in the partitioning will show where the problematic point is located Visipart checks the border of each subdomain in turn Because it has no information about boundary types Visipart cannot see the difference between closed and open boundaries Consequently it will produce more warnings than necessary If you are sure that a certain subdomain interface is OK select next segment in the dialog box to continue checking in the next part of the border of the subdomain Version 1 3 June 2011 9 User s Guide VISIPART Chapter 3 Optimizing grid partitionings The following approach has proven useful for the optimisation of a partitioning 10 1 An initial partitioning is created in the following steps a Read the computing domain s geometry using the function File Import WAQUA enclosure b Optionally select regions with the mouse to assign to subdomains with different numbers of layers for using vertical refinement c When using vertical refinement set the number of layers per subdomain using Options Set number of Layers d When using vertical refinement distribute the available processors over the subdomains using the function Options Assign processors to domain
5. 3 5 Edit Split subdomain A subdomain in the partition can easily be split up into smaller subdomains using one of the auto matic partitioning methods that are also available in Coppre The menu item Edit Split subdo main is used to split one subdomain into smaller subdomains After selecting this menu item a pointer cross hair comes up in the partitioning window which must be used to select a subdomain with the mouse Then a window pops up in which you can configure the automatic partitioning method to be used the number of subdomains that must be created the partitioning method to be used SGRB ORB or Strip wise the intial orientation for splitting the subdomain automatic rowwise or column wise The SGRB partitioning method is usually the best for partitioning of WAQUA grids It combines the favourable properties of the ORB and Strip wise methods It consists of recursively splitting the subdomain into two parts along the shortest dimension of each subdomain at each stage The ORB partitioning method consists of recursively splitting the subdomain into two parts thereby in each stage using an orientation orthogonal to the orientation of the previous stage The Strip wise partitioning method consists of splitting the subdomain into the requested number of parts using the same orientation for all subdomains When using the automatic orientation option the initial orientation depends on the sizes in m and n direc
6. a DDHOR simulation this must be done for each domain separately So first we load the first partition links and split it into 9 subdomains We end up with the assignment in Figure 14 Chapter 4 Example for assigning processes in DDHOR simulation using Visipart Visipart coppre r links10b modified File Edit View Options Help Figure 4 5 Domain links is now divided in 9 subdomains Next we load the second partition rebo and split it into 3 subdomains We end up with Figure Visipart coppre r rebol0b modified File Edit View Options Help 10 12 14 16 18 m Figure 4 6 Domain rebo is now divided in 3 subdomains And finally we open the third partition reon and split it into 4 subdomains so we end with Figure Version 1 3 June 2011 15 User s Guide VISIPART Visipart coppre r reon10b modified File Edit View Options Help Figure 4 7 Domain reon is now divided in 4 subdomains 16
7. be read and processed Then the current partitioning will be adjusted All grid points selected by the enclosure that have not been assigned to a subdomain will be assigned to a new subdomain npart 1 Then the subdomain value is kept for interior points and reset to O for other points of the grid Finally the resulting partitioning will be shown in the Visipart window Version 1 3 June 2011 3 User s Guide VISIPART 2 2 5 File Save partitioning To store a partitioning select this menu item This brings up a file selection dialogue in which the name of the file can be set The file is saved in the areas format using SUBDOMAIN s and ENCLOSURE s 2 2 6 File Print This menu item allows to print the current view of the partitioning using a printer or a file For this the Matlab standard printing options are used 2 2 7 File Exit This menu item is used to exit from Visipart When you have made modifications to the partitioning since the last moment you saved the partitioning or imported a new enclosure or partitioning a question will be asked whether you really want to exit or want to return to Visipart to save your work first 2 3 Options of the Edit menu 2 3 1 Edit Undo If you make an error while modifying the partitioning select this menu item to undo the latest modification 2 3 2 Edit Set subdomain number Left mouse button To modify a partitioning press and hold down the left mouse button at some poi
8. edure visipart pl This will start Matlab for you add the directory of the Visipart program files to the Matlab search directory and start its graphical interface This will bring up an empty figure window with a menu bar that is specific for Visipart Only one instance of Visipart may run at any moment Starting a second one results in an error message The perl procedure visipart pl is located in the bin directory of a SIMONA installation Visipart should work ok when the appropriate settings for the SIMONA installation have been made It is possible to start Visipart directly from the Matlab command prompt For this you should add a directory to the Matlab search path as follows addpath SIMONADIR matlab visipart After this Visipart can be started by executing the command visipart at the Matlab command prompt On Windows one can start Visipart by running the compiled version which can be found in lt SIMONADIR gt bin win One advantage of the compiled version is the possibility to have more than one instance of Visipart running at the same time To exit VISIPART select menu item File Exit 2 20 Options of the File menu 2 2 File Import WAQUA enclosure To start creating a new decomposition for a model geometry select an enclosure file via this menu item The enclosure is usually specified in the siminp file or an include file You may either select the whole siminp file to be used or select the include
9. el computing and areas files domain decomposition that are allowed by Coppre should be supported by Visipart too For information on their formats consult the Users guide for Parallel WAQUA TRIWAQ and for Domain decomposition Partitioning decomposition files may be created by hand by using Visipart see File Save partition ing or by using Coppre which writes the partitioning used to the report file coppre r lt lt runid gt Visipart determines the size of the grid that must be partitioned from the extension of the partitioning that is stored in the file So if you use an areas format and specify MNMNBOXes make sure that the extension of the boxes together match those of the grid itself Or otherwise start by reading the WAQUA enclosure into Visipart and after that apply the partitioning File Apply partitioning 2 2 4 File Apply WAQUA enclosure When you have read a partitioning as described above File Load partitioning it may not yet show the actual geometry of the WAQUA grid This may be accomplished by restricting the current par titioning to the grid points enclosed by a WAQUA computational grid enclosure This combination operation is activated by the current menu item When you select this menu item a pop up window shows up where you can select an enclosure file You may either select the whole siminp file to be used or select the include file that contains the computational grid enclosure The enclosure will
10. er subdomain the relative computing speeds are initially equal to 1 The actual relative computing speeds are set in a way very similar to the number of layers The relative computing speed can be any positive value larger than le 5 2 5 6 Options Show Observed Run times The relative computing speed per subdomain can be filled in automatically from the run times that are observed in an actual simulation using a given partitioning of the domain This option first brings up a file selection dialog box in which the waqpro message file should be selected This message file is then parsed by Visipart so it gets all the information needed Note that all lines are evaluated one by one so it is advised to keep the message file as short as possible For example we advise not to use the nmdbg application during a simulation when you want to use Visipart also Next another window is created in which the run times are displayed in a graph and estimates for the computing speed are shown in textual format The estimates are derived by combining the total calculation times and the workload number of grid points times number of layers of the actual decomposition The estimates are used in the estimated workload per domain only after the Accept speeds button is pressed Note that good estimates are obtained only when the message file corresponds to the current partitioning in Visipart In case of a DDHOR simulation the runtimes for all processors from each doma
11. file that contains the computational grid en closure The enclosure will be read processed all interior points will be assigned to subdomain 1 and the resulting partitioning will be shown in the Visipart window 2 Chapter 2 Description of the menu items of Visipart 2 2 2 File Apply partitioning When you have imported a WAQUA enclosure as described above File Import WAQUA enclo sure you may combine this with a partitioning that is described in a partitioning input file This is useful when the partitioning is given through possibly large rectangular boxes or enclosures that do not show the actual geometry When the partitioning is restricted to the actual grid this allows for reading curvilinear grid coordinates too The partitioning files supported here are the same as those supported by File Load partitioning described below This option can be used only when there have not been created any subdomains i e when all grid points of the enclosure are assigned to subdomain 1 Grid points that are active in the enclosure subdomain 1 and not active in the partitioning subdo main 0 will be assigned to a separate subdomain 2 2 3 File Load partitioning Select this menu item to load an existing partitioning This brings up a file selection dialogue from which the partitioning file can be selected Once the partitioning has successfully been loaded it will be shown in the Visipart window All partitioning files parall
12. he largest subdomain in a run scaled by the computing time for the run with the largest number of processors The computing cost is computed as the computing time times the number of processors used These values are scaled by the computing time for a single processor run This is equal to one over the efficiency of the parallel run Below the graph the optimal assignment is shown for one of the numbers of processors in the range When the decomposition of the global domain into regions is modified the computed distribution of processors over the regions may become sub optimal In this case the Recompute button may be pressed for determining the new optimal assignments of processors and for updating the graph For a DDHOR simulation this option is only available if the computing times for an actual run are available This is because Visipart only knows about one of the domains the measured run times provide information on the other domains If the observed runtimes are not available to Visipart yet Visipart asks whether this is a DDHOR simulation or not If you answer yes then Visipart will kindly ask you to load the observed run times first 2 5 8 Options Check Domain Decomposition This option can be used to check whether a partitioning is a valid domain decomposition parti tioning This option is enabled only when the grid is shown in rectilinear coordinates see also menu item View Show Curvilinear A partitioning is valid for
13. in are shown In the figure that pops up it is clearly shows which times are for which domain When the Accept speeds button is pressed a box pops up with the question which domain is currently being edited If an existing name is given and OK is pushed the estimated speeds for that domain are used for the estimated workload per domain See also Section 4 8 Chapter 2 Description of the menu items of Visipart 2 5 7 Options Assign Processors to Subdomains This option was developed with the vertical refinement option of Triwaq in mind particularly for the combined usage of domain decomposition and parallel computing The idea is that the domain decomposition is created first i e the regions with different numbers of layers are selected and then the different regions are further subdivided into an appropriate number of subdomains The option allows for computing an appropriate distribution of processors over the different regions thereby taking into account the workload in each region number of active grid points times number of layers When the option is selected this brings up a pop up window in which a range can be given for the number of processors that is considered Then for all numbers of processors in this range the optimal assignment of processors to regions is determined assuming perfect parallellisation and the resulting computing time and cost are set out in a graph The computing time shown is the estimated workload for t
14. mated workload Select this menu item for viewing an estimate of the relative amount of computing time needed per domain This estimate is derived from the number of grid points in each subdomain times the number of layers divided by the relative computing speed per domain when given see below This is a useful feature when creating a partitioning for parallel computing but will not often be used when selecting the subdomains for domain decomposition with vertical refinement 2 5 2 Options Show Fullbox This option is meant for giving information on the characteristics of each subdomain It shows the size of the rectangular hull around the domain which is the dimension of arrays the computational core of Waqua Triwaq 2 5 3 Options Show Fill ratio This option is meant for giving information on the characteristics of each subdomain It shows the ratio of the number of active wet grid points over the total number of grid points i e including land points in the full box 2 5 4 Options Set Number of Layers When using domain decomposition with vertical grid refinement DDVERT the number of layers may vary per subdomain This is relevant for Visipart when you are combining DDVERT with parallel computing when you want to split each DDVERT subdomain into different parts with about the same runtime per processor Version 1 3 June 2011 7 User s Guide VISIPART The number of layers per subdomain is used in estimating the workl
15. mple for assigning processes in DDHOR simulation using Visipart Visipart coppre r links10b Visipart coppre r rebol0b File Edit View Options Help File Edit View Options Help a domain links Visipart coppre r reonl0b File Edit View Options Help c domain reon Figure 4 1 The 3 domains of the example Visipart Observed run times wagpro m ddh BH total comm time MA other computing E calc transp y dir EH cal transp x dir MN calc Now y dir EB calc flow cir Figure 4 2 The observed runtimes of the example Version 1 3 June 2011 13 User s Guide VISIPART Figure 4 3 Question dialog display to select the current domain Now that the run times for all of the domain are known we can select Options Assign Processes to Subdomains This will work now in largely the same way as for a parallel run The next figure shows an optimal assignment for our example In our example we have 16 processors to work with but we do not use them all if the effect on the computing time is minimal see Figure 4 4 Assignment of processors to domains HE relative computing time T E relative cost P T Figure 4 4 Overview of how to assign processes over the domains From this figure we see that 9 of the processors should be used for domain links 3 for rebo and 4 for reon Now that the optimal assignment is known each domain can be split by using Edit Split sub domains In case of
16. nt in the partitioning and drag the mouse This will show a selected area in the partitioning Once you let go of the left mouse button a dialogue will appear asking for the subdomain number to which the selected area must be assigned Typing a number that does not yet occur in the partitioning will create a new subdomain Domains have a positive number but for one exception Subdomain 1 is used to identify the inactive part of the domain in a DDHOR simulation This inactive part should be filled in by another domain siminp file See also the example in Section 4 The menu item Edit Set subdomain number brings up a small informational message that points you to using the left mouse button 4 Chapter 2 Description of the menu items of Visipart 2 3 3 Edit Get subdomain number Right mouse button The subdomain numbers are visible in the partitioning In some cases it may be difficult to deter mine to which subdomain a particular grid point belongs Press the right mouse button at that grid point to get the number of the subdomain to which it belongs This is also useful for setting the default value for modifying areas with the left mouse button The menu item Edit Get subdomain number does nothing more than point you to using the right mouse button 2 3 4 Edit Remap To assign a new number to an entire subdomain select this menu item This will bring up a dialogue asking for the new subdomain number 2
17. oad per subdomain see Op tions Show Estimated Workload above and in suggesting the number of sub subdomains per subdomain in option Options Assign Processors to Subdomains below The number of layers per subdomain are initially all equal to 1 so that the estimated workload is equal to the number of active grid points per subdomain The actual numbers of layers in each subdomain can be entered in pop up windows that are brought up when this option is selected The pop up window displays the current values of the settings for all subdomains and the default value that is used when new subdomains are created These values can be modified by selecting an input region and typing the desired values The number of layers should be a positive integer value The new values are accepted when the pop up window is closed by pressing the OK button 2 5 5 Options Set Computing Speeds For each subdomain a relative computing speed may be entered This speed is used in constructing the graph with estimated workload per subdomain see Options Show Estimated Workload One application of using varying computing speeds per subdomain is when both faster and slower computers are used in a single run Another application is found in optimizing grid partitionings This latter application primarily uses Options Show Observed Run Times as described below and in Chapter on optimizing grid partitionings Like the number of layers p
18. s e Split the global domain when not using vertical refinement or each subdomain when using vertical refinement in the desired number of subdomains using Edit Split sub domain f Save the partitioning using File Save partitioning 2 Repeatedly carry out a short simulation approximately 100 time steps and improve the par tioning on basis of the actual computing times a Carry out a simulation using the current partitioning b Re start Visipart read the current partitioning and reenter the number of layers for each subdomain This step is not necessary if Matlab is still running c Read the actual CPU times from the waqpro message file using Options Show Ob served Runtimes If the observed runtimes seem reliable press Accept Speeds d Move computational grid points from one subdomain to another by selecting them with the mouse Keep moving points until the estimated workloads are as equal as possible e Save the partitioning using File Save partitioning Chapter 3 Optimizing grid partitionings Steps 1 a f and 2 a e provide the best standard procedure known for the optimisation of partition ings The following difficulties in the procedure are known 1 The assumptions made for the estimations of processor speeds on basis of run times The simulation calculations consist of several parts flow transport turbulence each with differ ent work load distributions It is possible that the proce
19. ssor with the heaviest work load is the restricting factor in none of these steps 2 The assumption that the processor speed does not change when the partitioning is changed Sometimes very small changes in the partitioning may have very large effects on the calcu lation speed mostly due to cache effects For example some computers work very slowly on arrays whose size is a multiple of 256 3 The procedure has to be restarted when the number of processors per domain i e the part of the global domain in which a certain number of layers has been selected is changed Refinements of the optimisation procedure and the support by Visipart may become necessary in the future Version 1 3 June 2011 11 User s Guide VISIPART Chapter 4 Example for assigning processes in DDHOR simulation using Visipart In this example we illustrate the use of Visipart for DDHORsimulations The first step for the user is to prepare the simulation input files for the different domains Next so called areas files must be devised for establishing the DDHOR coupling of the domains parts of the domain may be rendered inactive 1 the inactive areas must have matching openings or DDHORinterfaces This may be done using Visipart Thirdly a DDHOR config file is needed and an initial DDHORsimulation may be performed in which each DDHOR domain uses a single subdomain The way this file must be constructed is de scribed in Section 2 7 of the
20. tions of the initial subdomain If the subdomain has less rows than columns the initial split will be row wise assigning the first X rows to part 1 and the remaining Y rows to part 2 The implementation supports any number of subdomains For instance when you request 5 subdo mains the values of X and Y will be determined such that the amount of work for the first X rows is about 1 5 times the work of that for the remaining Y rows Then the method will recursively split the first part into 3 sub parts and the second part into 2 sub parts such that each part approximately gets Version 1 3 June 2011 5 User s Guide VISIPART the same amount of work Note that the estimated work load may be investigated using menu item Options Show workload 2 3 6 Edit Split all subdomains Menu item Edit Split all subdomains is used for splitting all subdomains simultaneously into smaller subdomains all using the same method as described under Edit Split subdomain above This option is in particular useful for creating a finer partition using an already existing optimised partition 2 4 Options of the View menu 2 4 1 View Zoom in To get a closer look at the partitioning select this menu item When only a part of the partitioning is in the figure sliders will be visible next to the partitioning These can be used to scroll through the partitioning The effect of the slider move will not become visible until de mo
21. use button is released 2 4 2 View Zoom out View a larger part of the partitioning The region that is displayed will be enlarged by approximately a factor 1 8 2 4 3 View Zoom Region When you have selected this menu item you will be prompted to mark the area in the grid that you want to have enlarged Push and hold down the left mouse button at one corner of the desired area which must be inside the figure showing the partitioning and drag the mouse to the opposite corner When you release the left mousebutton the zooming will take place 2 44 View Zoom Full Get a full view of the partitioning again 2 4 5 View Max Aspect Ratio This selection will rescale the figure to its maximal size in the Visipart window This option is not available when viewing a curvilinear grid 6 Chapter 2 Description of the menu items of Visipart 2 4 6 View Show Grid This option will display a coordinate grid over the partitioning to help determine the coordinates of points in the partitioning This option is only effective when viewing the grid using computational m n coordinates 2 4 7 View Show Curvilinear In case of a curvilinear grid the partitioning can be shown in physical coordinates by selecting this menu item When selecting this option for the first time for a particular grid you will be prompted for the name of the RGF file in which coordinates are stored 2 5 Options of the Options menu 2 5 1 Options Show esti
Download Pdf Manuals
Related Search