Parallel & Concurrent Programming: ZPL. Emery Berger CMPSCI 691W Spring 2006 AMHERST. Department of Computer Science UNIVERSITY OF MASSACHUSETTS

Transcription

1 Parallel & Concurrent Programming: ZPL Emery Berger CMPSCI 691W Spring 2006 Department of Computer Science

2 Outline Previously: MPI point-to-point & collective Complicated, far from problem abstraction OpenMP - parallel directives Today: Language extensions to Fortran/C/C++ Questionable semantics, error-prone Something way better: ZPL lecture material from ZPL project, UW Department of Computer Science 2

3 ZPL Parallel array language Implicitly parallel No parallel constructs per se Very high level Assignments work at array level, as in A := B + C Machine independent Compiles to ANSI C + communication library calls (e.g., MPI) Efficient Department of Computer Science 3

4 Comparison Matrix-multiplication: C triply-nested loop ZPL dot-product of rows & columns efficiently implemented on parallel machines Department of Computer Science 4

5 ZPL Outline Language overview Regions Directions Parallel array operations Handling boundary conditions ZPL programs & performance Department of Computer Science 5

6 Regions Key abstraction in ZPL: regions Index sets (rows,cols) partition matrices Operate on regions, not indexed items! rows columns Department of Computer Science 6

7 Region Examples Interior of matrix Left-most column Department of Computer Science 7

8 Directions Directions: Offset vectors used to manipulate regions & array data Department of Computer Science 8

9 Creating New Regions Prepositions create new regions: in of at by Applies direction to select part of region Creates new region outside existing region Shifts a region by a direction Creates new region strided by direction Department of Computer Science 9

10 Applying Directions Use in to apply direction to region + = Department of Computer Science 10

11 Create Region Outside Use of to create region outside existing region Extends region + = = Department of Computer Science 11

12 Shifting Regions Use at to create new region shifted by a direction + = Department of Computer Science 12

13 Striding Regions Use by to create new region strided by a direction Department of Computer Science 13

14 Parallel Arrays Parallel arrays declared over regions Department of Computer Science 14

15 Computing Over Arrays Can use regions as modifiers that define computations over arrays: Department of Computer Science 15

16 Arrays & Communication Most computations in ZPL do not involve communication Exceptions include: Shifting Reduction Broadcast All-to-all Department of Computer Science 16

17 Shifting operator shifts data in direction This translation induces point-to-point communication Department of Computer Science 17

18 Reduction Op<< computes reductions Reduction (tree-style) communication +<< (sum), *<< (times), min<<, max<< For prefix (scan), use op Department of Computer Science 18

19 Broadcast (Flooding) >> (flood) replicates data across dimensions of array Triggers broadcast operation Department of Computer Science 19

20 Mapping Remap (#) moves data between arrays Specified by map arrays Built-in Index1, Index2 Index1 = row indices, Index2 = col indices Department of Computer Science 20

21 Boundary Conditions Boundary conditions ( corner cases ) Usually tedious, error-prone Very simple in ZPL Department of Computer Science 21

22 Boundary Conditions Periodic boundary conditions with wrap Department of Computer Science 22

23 ZPL Example Jacobi iteration replace elements in array with average of four nearest neighbors, until largest change < δ Consider difficulty of parallelizing with MPI/OpenMP boundary conditions, etc Department of Computer Science 23

24 ZPL Example Department of Computer Science 24

25 ZPL Performance Department of Computer Science 25

26 ZPL Example: Life Department of Computer Science 26

27 ZPL Example: Life Department of Computer Science 27

28 The End Next time: Your turn! Occam & Multilisp Department of Computer Science 28