Beating Gridlock: Parallel Programming with SAS Grid Computing and SAS/CONNECT

Transcription

1 Beating Gridlock: Parallel Programming with SAS Grid Computing and SAS/CONNECT Presenter s Name Thursday, February 20, 2014 Jack Fuller Experis Business Analytics

2 Introduction The Problem A SAS program is taking a very long time to execute The program is constructed of multiple, independent subtasks Simulation programs frequently exhibit this behavior Experis Thursday, February 20,

3 Introduction The Solution Split the program into a driver program and a set of subprograms which run in parallel Elapsed Real Time Total CPU Time Experis Thursday, February 20,

4 Overview When to use parallel processing How to use parallel processing Points to consider Experis Thursday, February 20,

5 When to Use Parallel Processing

6 Subtasks Initialization Subtask 1 Subtask 2 Subtask 3 Finalization Experis Thursday, February 20,

7 Subtasks Initialization Subtask 1 Subtask 2 Subtask 3 Finalization Experis Thursday, February 20,

8 Subtasks Initialization Subtask 1 Subtask 2 Subtask 3 Finalization Experis Thursday, February 20,

9 Subtasks Initialization Subtask 1 Subtask 2 Subtask 3 Finalization Experis Thursday, February 20,

10 Subtasks Initialization Subtask 1 Subtask 2 Subtask 3 Finalization Experis Thursday, February 20,

11 Subtasks Initialization Subtask 1 Subtask 2 Subtask 3 Finalization Experis Thursday, February 20,

12 Subtasks SAS Connect 9.3 User s Guide Initialization Server Subtask 1 Subtask 2 Subtask 3 Client Client Client Finalization Server Experis Thursday, February 20,

13 Subtasks Grid Computing in SAS 9.3 Initialization Task Subtask 1 Subtask 2 Subtask 3 Subtask Subtask Subtask Finalization Task Experis Thursday, February 20,

14 Types of Concurrency Data Parallelism - subtasks work on different pieces of data Subtask 1 loads observations 1-10,000 Subtask 2 loads observations 10,001-20,000 Task Parallelism - subtasks perform different functions Subtask 1 processes demographics data Subtask 2 processes vital signs data Experis Thursday, February 20,

15 Amdahl s Law P is the number of processors is the fraction of work that is not parallelized Describes the optimal increase in speed gained from converting a serial process to a parallel process Experis Thursday, February 20,

16 Amdahl s Law as P is the number of processors is the fraction of work that is not parallelized As the number of processors (P) increases, the increase in speed is constrained by the fraction of work that is not parallelized (α) Experis Thursday, February 20,

17 Graphical Representation of Amdahl's Law Experis Thursday, February 20,

18 How to Use Parallel Processing

19 Our Problem Creating Test Data do pt=1 to 10000; if uniform(0) >.40 then do; gender = M ; /* Generate test data */ else do; gender = F ; /* Generate test data */ output; How do we want to define our concurrency? Experis Thursday, February 20,

20 Our Problem Creating Test Data do pt=1 to 10000; if uniform(0) >.40 then do; gender = M ; /* Generate test data */ else do; gender = F ; /* Generate test data */ output; Data Parallelism Experis Thursday, February 20,

21 Our Problem Creating Test Data do pt=1 to 10000; if uniform(0) >.40 then do; gender = M ; /* Generate test data */ else do; gender = F ; /* Generate test data */ output; Task Parallelism Experis Thursday, February 20,

22 Our Problem Creating Test Data do pt=1 to 10000; if uniform(0) >.40 then do; gender = M ; /* Generate test data */ else do; gender = F ; /* Generate test data */ output; Let s choose task parallelism by SDTM domain Experis Thursday, February 20,

23 Example Initialization %let seed = 0; %let numpt = 1000; Passed to each subtask %let rc=%sysfunc(grdsvc_enable(_all_, resource=sasapp)); options autosignon; libname shared "%sysfunc(pathname(work))"; Experis Thursday, February 20,

24 Example Initialization %let seed = 0; %let numpt = 1000; Enable the grid service %let rc=%sysfunc(grdsvc_enable(_all_, resource=sasapp)); options autosignon; libname shared "%sysfunc(pathname(work))"; Experis Thursday, February 20,

25 Example Initialization %let seed = 0; %let numpt = 1000; %let rc=%sysfunc(grdsvc_enable(_all_, resource=sasapp)); Allow the program to sign on without user interaction options autosignon; libname shared "%sysfunc(pathname(work))"; Experis Thursday, February 20,

26 Example Initialization %let seed = 0; %let numpt = 1000; %let rc=%sysfunc(grdsvc_enable(_all_, resource=sasapp)); Point a LIBREF to the task s work library options autosignon; libname shared "%sysfunc(pathname(work))"; Experis Thursday, February 20,

27 Example Create the Initial Data data PT pt; GENDER do 1 pt=1 to &numpt; 2 gender = ifc(uniform(&seed) >.45, 'M', 'F'); 3 output; run; Our input data lives in the task s work directory Experis Thursday, February 20,

28 Example Demographics Subtask %syslput seed=&seed / remote=task1; rsubmit task1 wait=no inheritlib=(shared); data shared.task1(keep=pt ht wt); set shared.pt; if gender = M then do; /* Compute statistics */ else do; /* Compute statistics */ run; endrsubmit; Create a macro variable in the subtask and initialize it to &seed Experis Thursday, February 20,

29 Example Demographics Subtask %syslput seed=&seed / remote=task1; rsubmit task1 wait=no inheritlib=(shared); data shared.task1(keep=pt ht wt); set shared.pt; if gender = M then do; /* Compute statistics */ else do; /* Compute statistics */ run; endrsubmit; Delineate the code which will run in the subtask Experis Thursday, February 20,

30 Example Demographics Subtask %syslput seed=&seed / remote=task1; rsubmit task1 wait=no inheritlib=(shared); data shared.task1(keep=pt ht wt); set shared.pt; if gender = M then do; /* Compute statistics */ else do; /* Compute statistics */ run; endrsubmit; Name of the subtask NOTE: Subtask names are limited to 8 chars Experis Thursday, February 20,

31 Example Demographics Subtask %syslput seed=&seed / remote=task1; rsubmit task1 wait=no inheritlib=(shared); data shared.task1(keep=pt ht wt); set shared.pt; if gender = M then do; /* Compute statistics */ else do; /* Compute statistics */ run; endrsubmit; Specify asynchronous execution Experis Thursday, February 20,

32 Example Demographics Subtask %syslput seed=&seed / remote=task1; rsubmit task1 wait=no inheritlib=(shared); data shared.task1(keep=pt ht wt); set shared.pt; if gender = M then do; /* Compute statistics */ else do; /* Compute statistics */ run; endrsubmit; Give the subtask access to the specified library Experis Thursday, February 20,

33 Example Demographics Subtask %syslput PT seed=&seed HT / remote=task1; WT rsubmit 1 task1 wait=no inheritlib=(shared); data 2 shared.task1(keep=pt ht wt); 3 set shared.pt; if gender = M then do; /* Compute statistics */ else do; /* Compute statistics */ run; endrsubmit; Experis Thursday, February 20,

34 Example Labs Subtask %syslput PT seed=&seed WBC / remote=task2; RBC rsubmit 1 task2 wait=no inheritlib=(shared); data 2 shared.task2(keep=pt wbc rbc); 3 set shared.pt; /* Compute statistics */ run; endrsubmit; Experis Thursday, February 20,

35 Example Vital Signs Subtask %syslput PT seed=&seed SYST_BP / remote=task3; DIAST_BP rsubmit 1 task3 wait=no inheritlib=(shared); data 2 shared.task3(keep=pt syst_bp diast_bp); 3 set shared.pt; if gender = M then do; /* Compute statistics */ else do; /* Compute statistics */ run; endrsubmit; Experis Thursday, February 20,

36 Example Finalization waitfor _all_ task1 task2 task3; data final; merge pt task:; by pt; run; signoff _all_; Wait for subtasks Experis Thursday, February 20,

37 Example Finalization waitfor PT GENDER _all_ HT task1 WT task2 WBC task3; RBC SYST_BP DIAST_BP 1 data 2 final; 3 merge pt task:; by pt; run; signoff _all_; Notice the use of the task work library Experis Thursday, February 20,

38 Example Finalization waitfor _all_ task1 task2 task3; data final; merge pt task:; by pt; run; signoff _all_; Sign off all subtasks Experis Thursday, February 20,

39 Points to Consider

40 How many subtasks should there be? This will depend on a number of factors The fraction of work that is not parallelized The number of available grid nodes My rule of thumb I like to benchmark my applications at 5, 10 and 15 subtasks and then adjust accordingly Experis Thursday, February 20,

41 How much work should occur in each subtask? There has to be enough work in each subtask SAS/CONNECT has a login process which I have benchmarked as taking between 5 and 15 seconds The goal is to have the friction involved with creating/managing the subtasks become negligible My rule of thumb I like to have each subtask take at around 10 minutes Experis Thursday, February 20,

42 Should libraries be shared between subtasks? Arguments for using INHERITLIB One thing is in one place Arguments against using INHERITLIB Subtasks can bottleneck as they queue up awaiting access My rule of thumb I like to use INHERITLIB unless I have a reason not to Experis Thursday, February 20,

43 Conclusion

44 Conclusion Put as much work as possible in the subtasks Get something up and running, benchmark it and then adjust things Good luck! Experis Thursday, February 20,

45 Questions

46 Acknowledgement SAS and all other SAS Institute Inc. product or service names are registered trademarks or trademarks of SAS Institute Inc. in the USA and other countries. indicates USA registration. Other brand and product names are registered trademarks or trademarks of their respective companies. Experis Thursday, February 20,

47 Thank you Jack Fuller Senior Application Developer Experis Thursday, February 20,