automatically exploiting cross-invocation parallelism using runtime ...

More documents

Recommendations

Info

for a variety of reasons. For instance, each thread may be assigned different number oftasks and the execution time of each task may vary. Threads that finish the inner loop earlymay not execute past the barrier, resulting in very poor scalability. Experimental results inFigure 3.3 show that performance is worse than the sequential version and worsens as thenumber of threads increases.Figure 3.2(b) shows the execution plan after DOMORE’s partitioning phase (Section3.3.1). The first thread executes code in the outer loop (statement A to D) and servesas the scheduler. The other threads execute update code in the inner loop concurrentlyand serve as workers. Overlapping the execution of scheduler and worker threads improvesthe performance, however, without enabling cross-invocation parallelism, clock cycles arestill wasted at the synchronization points.Figure 3.2(c) shows the execution plan after the DOMORE transformation completesand enables cross-invocation parallelization. The scheduler sends synchronization informationto the worker threads and worker threads only stall when a dynamic dependenceis detected. In the example, most of CG’s iterations are independent and may run concurrentlywithout synchronization. However, iteration 1.5 (i.e. when i=1, j=5) updates memorylocations which are later accessed by iteration 2.2. At runtime, the scheduler discovers thisdependence and signals thread one to wait for iteration 1.5. After synchronization, threadone proceeds to iteration 2.2. As shown in Figure 3.3, DOMORE enables scalable loopspeedup for CG up to 24 threads on an 24-core machine.Figure 3.4 shows a high-level overview of the DOMORE transformation and runtimesynchronization scheme. DOMORE accepts a sequential program as input and targets hotloop nests within the program. At compile-time, DOMORE first partitions the outer loopinto a scheduler thread and several worker threads (Section 3.3.1). The scheduler threadcontains the sequential outer loop while worker threads contain the parallel inner loop.Based on the parallelization technique used to parallelize the inner loop, the code generationalgorithm generates a multi-threaded program with cross-invocation parallelization26
thread1 thread2 thread3 thread4thread1 thread2 thread3 thread4thread1thread2 thread3 thread4A1B1C1D1.1E1.1D1.5E1.5A2B2C2D2.1D1.2E1.2D2.2BarrierD1.3E1.3Barrier PenaltyIdle CoresD2.3D1.4E1.4D2.4A1B1C1D1.1D1.2D1.3D1.4D1.5A2B2C2D2.1D2.2D2.3D2.4D2.5E1.1E1.4E2.1E1.2E1.5BarrierE2.2E1.3BarrierPenaltyE2.3A1B1C1D1.1ScheduleD1.2ScheduleD1.3ScheduleD1.4ScheduleD1.5ScheduleA2B2C2D2.1ScheduleD2.2ScheduleD2.3ScheduleD2.4ScheduleD2.5ScheduleE1.1E1.2E1.4E1.5stallE2.2E2.3E1.3E2.1E2.1E2.2E2.3E2.4E2.5E2.5E2.4D2.5E2.5Barrier PenaltyIdle CoresE2.4BarrierPenaltyDOMOREfinalDOMOREafter partitioningDOALL(a)(b)(c)Figure 3.2: Comparison of performance with and without cross-invocation parallelization :(a) DOALL is applied to the inner loop. Frequent barrier synchronization occurs betweenthe boundary of the inner and outer loops. (b) After the partitioning phase, DOMORE haspartitioned the code without inserting the runtime engine. A scheduler and three workersexecute concurrently, but worker threads still synchronize after each invocation. (c) DO-MORE finalizes by inserting the runtime engine to exploit cross-invocation parallelism.Assuming iteration 2 from invocation 2 (2.2) depends on iteration 5 from invocation 1(1.5). Scheduler detects the dependence and synchronizes those two iterations.(Section 3.3.5). At runtime, the scheduler thread checks for dynamic dependences, schedulesinner loop iterations, and forwards synchronization conditions to worker threads (Section4.2). Worker threads use the synchronization conditions to determine when they areready to execute.27
Page 1 and 2: AUTOMATICALLY EXPLOITINGCROSS-INVOC
Page 4 and 5: techniques. Among twenty programs f
Page 6 and 7: in particular. Their professionalis
Page 8 and 9: ContentsAbstract . . . . . . . . .
Page 10 and 11: 4.5.4 Load Balancing Techniques . .
Page 12 and 13: 2.4 Sequential Loop Example for DOA
Page 14 and 15: 4.5 Overview of SPECCROSS: At compi
Page 16 and 17: 1.1 Limitations of Existing Approac
Page 18 and 19: advanced forms of parallelism (MPI,
Page 20 and 21: the graph stands for an iteration i
Page 22 and 23: 1.2 ContributionsFigure 1.5 demonst
Page 24 and 25: 1.3 Dissertation OrganizationChapte
Page 26 and 27: alias the array regular via inter-p
Page 28 and 29: 1 for (i = 0; i < M; i++){2 node =
Page 30 and 31: 1 cost = 0;2 node = list->head;3 Wh
Page 32 and 33: example which cannot benefit from e
Page 34 and 35: These techniques are referred to as
Page 36 and 37: unnecessary overhead at runtime.Tab
Page 38 and 39: Chapter 3Non-Speculatively Exploiti
Page 42 and 43: 12x11x10xDOMOREPthread Barrier9xLoo
Page 44 and 45: Algorithm 1: Pseudo-code for schedu
Page 46 and 47: Algorithm 2: Pseudo-code for worker
Page 48 and 49: 3.3 Compiler ImplementationThe DOMO
Page 50 and 51: to T i . DOMORE’s MTCG follows th
Page 52 and 53: Outer_Preheaderbr BB1ABB1A:ind1 = P
Page 54 and 55: Algorithm 3: Pseudo-code for genera
Page 56 and 57: Scheduler Function SchedulerSync Fu
Page 58 and 59: SchedulerWorker1 Worker2Worker3Work
Page 60 and 61: 3.5 Related Work3.5.1 Cross-invocat
Page 62 and 63: ations during the inspecting proces
Page 64 and 65: for (t = 0; t < STEP; t++) {L1: for
Page 66 and 67: sequential_func() {for (t = 0; t <
Page 68 and 69: Workerthread 1Workerthread 2Workert
Page 70 and 71: library provides efficient misspecu
Page 72 and 73: Workerthread 1TimeFigure 4.6: Timin
Page 74 and 75: 4.2 SPECCROSS Runtime System4.2.1 M
Page 76 and 77: takes up to 200MB memory space.To d
Page 78 and 79: checkpoint, the child spawns new wo
Page 80 and 81: Operation DescriptionFunctions for
Page 82 and 83: Main thread:main() {init();create_t
Page 84 and 85: implemented in the Liberty parallel
Page 86 and 87: Algorithm 5: Pseudo-code for SPECCR
Page 88 and 89: CROSS, since SPECCROSS can be appli
Page 90 and 91:
techniques.Synchronization via sche
Page 92 and 93:
Source Benchmark Function % of exec
Page 94 and 95:
applied to the outermost loop, gene
Page 96 and 97:
5.2 SPECCROSS Performance Evaluatio
Page 98 and 99:
and the number of checking requests
Page 100 and 101:
8x7xno misspec.with misspec.Geomean
Page 102 and 103:
This thesisPrevious workSpeedup (x)
Page 104 and 105:
Program Speedup6x5x4x3x2xLOCALWRITE
Page 106 and 107:
for DOMORE and SPECCROSS. Others (e
Page 108 and 109:
programs and it achieves a geomean
Page 110 and 111:
Bibliography[1] R. Allen and K. Ken
Page 112 and 113:
[15] R. Cytron. DOACROSS: Beyond ve
Page 114 and 115:
[31] T. B. Jablin, Y. Zhang, J. A.
Page 116 and 117:
[47] A. Nicolau, G. Li, A. V. Veide
Page 118 and 119:
[62] L. Rauchwerger and D. Padua. T
Page 120:
national conference on Parallel Arc
show all

automatically exploiting cross-invocation parallelism using runtime ...

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?