Mercurial > repos > public > sbplib_julia
annotate distributedTest.jl @ 149:11b6646918d4 parallel_test
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
author | Vidar Stiernström <vidar.stiernstrom@it.uu.se> |
---|---|
date | Tue, 26 Feb 2019 11:09:30 +0100 |
parents | 95a3ba70bccb |
children | 4dc19757cada |
rev | line source |
---|---|
148
95a3ba70bccb
Added some clarifying comments on the state of distributedTest.jl
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
147
diff
changeset
|
1 #TODO: everywhere using here or just everywhere include? |
146
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
2 @everywhere using DistributedArrays |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
3 |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
4 # TODO: Currently uses integer division to calculate the local grid size. |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
5 # Should we make sure this is handled in some way if mod(sz./nworkers()) != 0 |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
6 # or keep assertions? |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
7 @everywhere function create_partitioned_grid(size::NTuple{Dim, Int}, limit_lower::NTuple{Dim, T}, limit_upper::NTuple{Dim, T}, nworkers_per_dim::NTuple{Dim, Int}) where Dim where T |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
8 @assert mod.(size, nworkers_per_dim) == (0,0) |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
9 @assert prod(nworkers_per_dim) == nworkers() |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
10 # Translate the current worker id to a cartesian index, based on nworkers_per_dim |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
11 ci = CartesianIndices(nworkers_per_dim); |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
12 id = Tuple(ci[myid()-1]) |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
13 |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
14 # Compute the size of each partitioned grid |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
15 size_partition = size./nworkers_per_dim |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
16 size_partition = map(x->Int(x),size_partition) # TODO: Cant this be done in an easier way?... |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
17 # Compute domain size for each partition |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
18 domain_size = limit_upper.-limit_lower |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
19 domain_partition_size = domain_size./nworkers_per_dim |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
20 |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
21 # Compute the lower and upper limit for each grid partition, then construct the grid |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
22 ll_partition = limit_lower .+ domain_partition_size.*(id.-1) |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
23 lu_partition = limit_lower .+ domain_partition_size.*id |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
24 grid = sbp.Grid.EquidistantGrid(size_partition, ll_partition, lu_partition) |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
25 return grid |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
26 end |
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
27 |
149
11b6646918d4
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
148
diff
changeset
|
28 @everywhere function timed_apply(op, u, v) |
11b6646918d4
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
148
diff
changeset
|
29 @time sbp.apply_tiled!(op, u, v) |
11b6646918d4
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
148
diff
changeset
|
30 return nothing |
11b6646918d4
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
148
diff
changeset
|
31 end |
148
95a3ba70bccb
Added some clarifying comments on the state of distributedTest.jl
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
147
diff
changeset
|
32 |
149
11b6646918d4
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
148
diff
changeset
|
33 gridsize = (10000, 10000); # Global grid size |
148
95a3ba70bccb
Added some clarifying comments on the state of distributedTest.jl
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
147
diff
changeset
|
34 v = dzeros(gridsize) # Distribured arrays |
95a3ba70bccb
Added some clarifying comments on the state of distributedTest.jl
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
147
diff
changeset
|
35 u = dzeros(gridsize) # Distribured arrays |
146
21b188f38358
Started branch for testing out parallel computing
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
diff
changeset
|
36 |
149
11b6646918d4
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
148
diff
changeset
|
37 @sync @distributed for p in workers() |
11b6646918d4
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
148
diff
changeset
|
38 #Should these be declared globally or locally? |
11b6646918d4
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
148
diff
changeset
|
39 limit_lower = (0., 0.) |
11b6646918d4
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
148
diff
changeset
|
40 limit_upper = (2pi, 3pi/2) |
11b6646918d4
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
148
diff
changeset
|
41 nworkers_per_dim = (Int(nworkers()/2),Int(nworkers()/2)) |
11b6646918d4
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
148
diff
changeset
|
42 init(x,y) = sin(x) + sin(y) |
11b6646918d4
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
148
diff
changeset
|
43 grid = create_partitioned_grid(gridsize, limit_lower , limit_upper, nworkers_per_dim) |
11b6646918d4
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
148
diff
changeset
|
44 @inbounds v[:L] = sbp.Grid.evalOn(grid, init) |
11b6646918d4
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
148
diff
changeset
|
45 op = sbp.readOperator("d2_4th.txt","h_4th.txt") |
11b6646918d4
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
148
diff
changeset
|
46 Δ = sbp.Laplace(grid, 1.0, op) |
11b6646918d4
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
148
diff
changeset
|
47 @inbounds timed_apply(Δ,u[:L], v[:L]) |
11b6646918d4
Rewrote the test using the @distributed macro instead of spawns. Seems to improve results.
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
148
diff
changeset
|
48 end |
147
aa18b7bf4926
Changed to initialize u and v as distributed arrays
Vidar Stiernström <vidar.stiernstrom@it.uu.se>
parents:
146
diff
changeset
|
49 @show maximum(abs.(u + v)) |