Home

downloading

1. WAIT HELD DRMAA PS USER ON HOLD WAIT TIME DRMAA PS QUEUED ACTIVE WAIT LICENSES DRMAA PS QUEUED ACTIVE WAIT ASSOC JOB LIMIT DRMAA PS QUEUED ACTIVE WAIT ASSOC RESOURCE LIMIT DRMAA PS QUEUED ACTIVE WAIT ASSOC TIME LIMIT DRMAA PS QUEUED ACTIVE WAIT RESERVATION DRMAA PS QUEUED ACTIVE WAIT NODE NOT AVAIL DRMAA PS QUEUED ACTIVE WAIT TBD1 DRMAA PS QUEUED ACTIVE WAIT TBD2 DRMAA PS QUEUED ACTIVE FAIL DOWN PARTITION DRMAA PS FAILED FAIL DOWN NODE DRMAA PS FAILED FAIL BAD CONSTRAINTS DRMAA PS FAILED FAIL SYSTEM DRMAA PS FAILED FAIL LAUNCH DRMAA PS FAILED FAIL EXIT CODE DRMAA PS FAILED FAIL TIMEOUT DRMAA PS FAILED FAIL INACTIVE LIMIT DRMAA PS FAILED FAIL BANK ACCOUNT DRMAA PS FAILED For value greater than 0x4000 we do not update DRMAA state because describes a transient state 6 3 3 drmaa control operations mapping DRMAA LoadLeveler Comments DRMAA_CONTROL_SUSPEND slurm_suspend Administrators only DRMAA_CONTROL_RESUME slurm_resume Administrators only DRMAA_CONTROL_HOLD priority 0 DRMAA_CONTROL_RELEASE priority 0 Administrators only on hold save job priority and on release bring old value When job is submited in held state after resume it gets MAX UINT32_T value D
2. triggered mode passed official OGF DRMAA testsuite on Torque and PBS Pro systems 5 Improved DRMAA Scalability program A dedicated test program was written for scalability testing purposes It simulates SMOA Computing service under continious workload Addition thread killing jobs was added This program takes 5 arguments SUB_INT SLEEP_TIME POLL_INTERVAL WAIT_INTERVAL KILL_RANGE and starts 4 threads e first thread submits sleep job with SLEEP_TIME parameter e second thread every POOL_INTERVAL checks status of every job e third thread with WAIT INTERVAL timeout and SESSION ANY job parameters runs drmaa_wait and updates finished jobs list e fourth thread kills a random job every k seconds where ks is chosen randomly from interval O KILL RANGE Jobs are submited until user press ctrl c Then first second and fourth thread stops and program waits for every submitted job end 6 DRMAA implementation for SLURM 6 1 Introduction PSNC DRMAA for Simple Linux Utility for Resource Management SLURM is an implementation of Open Grid Forum DRMAA 1 0 Distributed Resource Management Application API specification for submission and control of jobs to Simple Linux Utility for Resource Management SLURM Using DRMAA grid applications builders portal developers and ISVs can use the same high level API to link their software with different cluster resource management systems This software also enables the integration of SMOA Com
3. Poznan Supercomputing and Networking Center Summer Projects Micha Mat oka michal matloka student put poznan pl Date September 23 2010 Poznan Contents 1 2 Automated code review of SMOA Services 1 2 1 Flawfnder 2 Leo A l 2 2 Rough Auditing Tool for Security RATS 2 plint 2 3 Integration with PL grid accounting infrastructure bat_updater module 3 4 PBS DRMAA implementing triggered mode 4 5 Improved DRMAA Scalability program 4 6 DRMAA implementation for SLURM 5 XR oar a Aa a PORE SS A PUE X ERE Ge GS REE wik 5 6 2 Input Documents x Va xx aa a 5 6 Mapping of DRMAA interface on SLURM API 2 0 2 2 ee 5 6 3 1 DRMAA job attributes mapping to aiu eI D 5 SUYTITUPTT PTT 6 6 5 drmaa control operations mapping LL ee T7 6 4 Native specification lt srias sco wcu s e e wo ei eau Wie ee a he b e R p T 6 DRMAA library implementation using DRMAA Utils 2 2 2 8 nd User Man Bl xox ch ae aR RES B3 OY WXe Re me 63 3X E63 4g 8 C ee eee oe ee i ee 8 1 Introduction During my summer work in Poznan Supercomputing Networking Center have worked on a few projects where the most important are listed below I Automated code review of SMOA Services 2 Integration with PL grid accounting infrastructure bat updater module 3 PBS DRMAA implementing triggered mode 4 DRMAA implementation for SLURM Total time spent on projects was one mo
4. RMAA_CONTROL_TERMINATE slumr_kill_job job_id SIGKILL 0 6 4 Native specification DRMAA interface allows to pass DRM dependent job submission options Those options may be specified directly by setting drmaa_native_specification job template attribute or indirectly by the drmaa_job_category job template attribute In SLURM DRMAA the following set of sbatch arguments were implemented Native specification Description A account name Charge job to specified accounts acctg freq Define the job accounting sampling interval comment An arbitrary comment C constraint list Specify a list of constraints contiguous If set then the allocated nodes must form a contiguous set exclusive Allocate nodenumber of tasks to invoke on each nodes in exclusive mode when cpu consumable resource is enabled mem MB Minimum amount of real memory mem per cpu MB Maximum mount of real memory per allocated cpu required by a job mincpus n Minimum number of logical processors threads per node N nodes N Number of nodes on which to run N min max ntasks per node n Number of tasks to invoke on each node p partition partition Partition requested qos qos Quality of Service requeue If set permit the job to be requeued reservation name Allocate resources from named reservation s share Job allocation can share nodes with other running jobs w nodelist hosts Request a specific l
5. id gt lt job gt lt batch_server gt mich laptop lt batch_server gt lt job_id gt 10732 mich laptop lt job_id gt lt user gt mich lt user gt lt group gt mich lt group gt lt queue gt bat ch lt queue gt lt ctime gt 1285112777 lt ctime gt lt qtime gt 1285112777 lt qtime gt lt etime gt 1285112777 lt etime gt lt start gt 16518 lt start gt lt end gt 1285112868 lt end gt lt exec_host gt lt node gt lt nodename gt mich laptop lt nodename gt lt cpu gt 0 lt cpu gt lt node gt lt exec_host gt lt cputime gt 0 lt cputime gt lt walltime gt 91 lt walltime gt lt mem gt 2547712 lt mem gt lt vmem gt 29581312 lt vmem gt lt estatus gt 0 lt estatus gt lt infrastructure gt smoa lt infrastructure gt lt grid_job_id gt 368e7fb7 666f 467c ab84 820804a0372d lt grid_job_id gt lt userDN gt anonymous userDN job lt site gt In this module a C ActiveMQ implementation is used because C version is not fully functional A C wrapper interface composed of three functions activemqcpp connect activemqcpp send message activemqcpp disconnect was written in order to allow communication with ActiveMQ from bat updater module 4 PBS DRMAA implementing triggered mode Triggered mode allows DRMAA to acquire job state change events This mode may be the only solution for many production clusters e g reef man poznan pl wait thread pbs home mode keep comple
6. ipc c 227 1 buffer read Check buffer boundaries if used in a loop Program output had 1858 lines Flawfinder displays mainly tips for programmers what they should be aware of while using certain functions Eventually all verified by me warnings were were not dangerous 2 2 Rough Auditing Tool for Security RATS Output fragment Total lines analyzed 309 Total time 0 000270 seconds 1144444 lines per second Entries in perl database 33 Entries in ruby database 46 Entries in python database 62 Entries in c database 334 Entries in php database 55 Analyzing ipc c ipc c 52 High fixed size local buffer ipc c 123 High fixed size local buffer ipc c 128 High fixed size local buffer Extra care should be taken to ensure that character arrays that are allocated on the stack are used safely They are prime targets for buffer overflow attacks Program output had 593 791 lines depends on check level Like in Flawfinder user gets list of tips and in all checked cases they were not dangerous in SMOA Core 2 3 Splint Output fragment ipc c in function sm ipc send fd ipc c 85 5 Arrow access of non pointer struct iovec 1 iov iov base Types are incompatible Use type to inhibit warning ipc c 86 5 Arrow access of non pointer struct iovec 1 iov iov len ipc c in function sm ipc recv fd ipc c 133 126 Function assert expects arg 1 to be boolean gets int fd received ipc c 157 5 Arrow access of non p
7. ist of hosts Additional description of each parameter can be found in man sbatch 6 5 DRMAA library implementation using DRMAA Utils h i http gforge man poznan pl at pro ect snoaincubator scnsvm acts ion NUITS VETUS TUE ut ils and tools like 6 6 End User Manual Documentation can be found in doc folder in the library package 6 7 DRMAA testsuite Library covers all DRMAA 1 0 specification with exceptions listed below It was successfully tested with Simple Linux Utility for Resource Management SLURM 2 1 13 on Linux and passes 44 44 tests of the official DRMAA test suite and is therefore DRMAA1 0 compliant for test suite version 1 7 2 Using non administrator account SLURM DRMAA passes 40 44 tests of the official DRMAA test suite Known limitations e fuction drmaa_control with DRMAA_CONTROL_HOLD DRMAA_CONTROL_RELEASE DRMAA_CONTROL_SUSPEND DRMAA_CONTROL_RESUME is administrator only e drmaa_wct_slimit not implemented e optional attributes drmaa deadline time drmaa duration hlimit drmaa duration slimit drmaa transfer files not implemented
8. nth 2 Automated code review of SMOA Services Automated code review software checks source code in order to find violations of program standards and displays them as the list of warnings My task was to use these tools to check the SMOA Core library 2 1 Flawfinder Output fragment Hits 6 Lines analyzed 308 in 0 52 seconds 18364 lines second Physical Source Lines of Code SLOC 225 Hits level 0 0 1 3 2 2 3 0 4 1 5 0 Hits level 0 6 14 6 24 3 3 1 4 1 5 0 Hits KSLOC level 0 26 6667 1 26 6667 2 13 3333 3 4 44444 4 4 44444 5 0 Minimum risk level 1 Not every hit is necessarily a security vulnerability There may be other security vulnerabilities review your code Flawfinder version 1 27 C 2001 2004 David A Wheeler Number of dangerous functions in C C ruleset 160 Examining ipc c ipc c 52 2 buffer char Statically sized arrays can be overflowed Perform bounds checking use functions that limit length or ensure that the size is larger than the maximum possible length ipc c 123 2 buffer char Statically sized arrays can be overflowed Perform bounds checking use functions that limit length or ensure that the size is larger than the maximum possible length ipc c 128 2 buffer char Statically sized arrays can be overflowed Perform bounds checking use functions that limit length or ensure that the size is larger than the maximum possible length
9. ointer struct iovec 1 iov gt iov_base ipc c 158 5 Arrow access of non pointer struct iovec 1 iov iov len ipc c in function sm simple read ipc c 220 110 Function assert expects arg 1 to be boolean gets void buf ipc c in function sm_simple_write ipc c 256 110 Function assert expects arg 1 to be boolean gets void buf ipc c 39 19 File static variable rcsid declared but not used A variable is declared but never used Program output had 2909 lines Splint is a more strict software It does not display tips but warnings concerning given source code Program output had 2909 lines have verified 1O most frequently occuring warnings and all of them was not dangerous in SM0A Core Example uuid c 111 3 Format argument 4 to snprintf 02x expects unsigned int gets int int uuid gt time_low 3 snprintf set to get integer but got size t odbc c 537 2 Assignment of unsigned short int to int posix time tm mon timestamp gt month 1 3 Integration with PL grid accounting infrastructure bat updater module bat updater is a SMOA Computing module which allows publication of resource usage data via DR MAA Transport layer is based on JMS Java Messaging Service and uses opensource implementa tion of ActiveMQ called Apache ActiveMQ CPP qu cc Iu Obtained rusage data are converted to XML format Example genarated XML via DRMAA interface lt xml version 1 0 encoding UTF 8 gt site name psnc smoa plgr
10. ommand char script contains generated bash script drmaa_start_time time_t begin_time drmaa_transfer_files OPTIONAL ATTRIBUTE NOT IMPLEMENTED practically not used because every cluster has shared file system drmaa_v_argv Add to bash script drmaa_v_email char mail_user drmaa_v_env char environment uint32_t env_size name value pairs one per line drmaa_wct_hlimit uint32_t time_limit drmaa_wct_slimit NOT IMPLEMENTED drmaa_wd char work_dir 6 3 2 DRMAA states mapping The DRMAA states list was compared with SLURM states that can be retrieved using API SLURM DRMAA JOB PENDING in next table JOB RUNNING DRMAA PS RUNNING JOB SUSPENDED DRMAA PS USER SUSPENDED JOB COMPLETE DRMAA PS DONE JOB CANCELLED DRMAA PS FAILED JOB FAILED DRMAA PS FAILED JOB TIMEOUT DRMAA PS FAILED JOB NODE FAIL DRMAA PS FAILED JOB PENDING is additionaly described by state reason variable SLURM DRMAA WATT_NO_REASON DRMAA_PS_QUEUED_ACTIVE WATT_PRIORITY DRMAA_PS_QUEUED_ACTIVE WAIT DEPENDENCY DRMAA PS QUEUED ACTIVE WAIT RESOURCES DRMAA PS QUEUED ACTIVE WAIT PART NODE LIMIT DRMAA PS QUEUED ACTIVE WAIT PART TIME LIMIT DRMAA PS QUEUED ACTIVE WAIT PART STATE DRMAA PS QUEUED ACTIVE
11. puting with the underlying LoadLeveler system for remote multi user job submission and control over Web Services 6 2 Input Documents e DRMAA 1 0 Grid Recommendation http www ogf org documents GFD 133 pdf e DRMAA C Binding v1 0 https forge gridforum org sf docman do downloadDocument projects drmaa wg docman root ggf_13 doc5545 e Simple Linux Utility for Resource Management SLURM Documentation https computing llnl gov linux slurm documentation html 6 3 Mapping of DRMAA interface on SLURM API 6 3 1 DRMAA job attributes mapping to SLURM SLURM job is described by job desc msg t structure DRMAA SLURM Comment drmaa_block_email do not set mail_user drmaa_deadline_time OPTIONAL ATTRIBUTE NOT IMPLEMENTED drmaa_duration_hlimit OPTIONAL ATTRIBUTE NOT IMPLEMENTED drmaa_duration_slimit OPTIONAL ATTRIBUTE NOT IMPLEMENTED drmaa_error_path char std_err drmaa_input_path char std_in drmaa_job_category library configuration Slurm_drmaa conf use native specification function Syntax like in native specification attribute account My_job N 1 2 drmaa_job_name char name drmaa_join_files same values of std_err and std_out drmaa_js_state uint32_t priority O held drmaa_native_specification arguments like in sbatch Syntax example account My_job N 1 2 drmaa_output_path char std_out drmaa_remote_c
12. ted needed comments 0 not set polling yes default configuration not set polling yes more effective than above I set triggered no read access to server logs needed Configuration polling and triggered modes At the start of the DRMAA program the current log file s size is stored Later in the wait thread log file parsing starts from the previously remebered position Log file name is determined by pbs_home configuration variable Path to Torque PBS Pro spool directory that contains server logs e g var spool pbs This way only new lines releted to current execution are being parsed On day change old file is being parsed once more and then new path to log file is generated Every Log file line is composed of 6 fields FLD DATE FLD EVENT FLD_OBJ FLD_TYPE FLD_ID FLD_MSG FLD EVENT FLD MSG comments 0008 Job Run at request of On Job run we launch job_ps in order to get execution host etc 0008 Job Modified at request of From FLD_DATE we get modify Scheduler mich laptop time 0008 Job Queued at request of owner job name queue 0008 Job deleted at request of 0010 Exit_status resources_used cput Job completed status and resource resources_used mem usage information parse resources_used vmem resources used walltime Every line gives information about job state Collected data in triggered mode is passed to pbsdrmaa job update status routine PBS DRMAA in

downloading

Contents

Download Pdf Manuals

Related Search

Related Contents