Add operations/arguments to local CuPy array benchmark #524
Conversation
Codecov Report
@@ Coverage Diff @@
## branch-0.18 #524 +/- ##
===============================================
+ Coverage 90.42% 92.45% +2.02%
===============================================
Files 15 16 +1
Lines 1128 1550 +422
===============================================
+ Hits 1020 1433 +413
- Misses 108 117 +9
Continue to review full report at Codecov.
|
madsbk
added
3 - Ready for Review
There was a problem hiding this comment.
Looks good to me, nice addition @charlesbluca.
Should there be an additional argument to specify the array slicing interval (which is currently fixed at 3)?
I think fixing it to 3 is fine for now. If we need to benchmark the impact of step sizes, we can always add it. At that point, we might consider using the first position argument as a command and use the following arguments as command-specific (e.g. like how git works)
Could the JSON output be cleaned up? Currently, a (truncated) sample output file looks like:
Looks fine to me :)
dask_cuda/benchmarks/local_cupy.py
Outdated
| @@ -179,6 +197,40 @@ async def run(args): | |||
| ) | |||
| print(fmt % (d1, d2, bw[0], bw[1], bw[2], total_nbytes[(d1, d2)])) | |||
|
|
|||
| if args.benchmark_json: | |||
| import json | |||
There was a problem hiding this comment.
Moving import json to the top should be fine
There was a problem hiding this comment.
Done! Also going to change this to from json import dump.
| x = rs.random((args.size, args.size), chunks=args.chunk_size).persist() | ||
| await wait(x) | ||
| func_args = (x,) | ||
|
|
pentschev
added
5 - Ready to Merge
|
@gpucibot merge |
This PR adds the following operations to the local CuPy array benchmark:
This also adds an additional special argument,
--benchmark-json, which takes an optional path to dump the results of the benchmark in JSON format. This would allow us to generate plots using the output, as discussed in #517.Some thoughts:
{ "operation": "transpose_sum", "size": 10000, "second_size": 1000, "chunk_size": 2500, "compute_size": [ 10000, 10000 ], "compute_chunk_size": [ 2500, 2500 ], "ignore_size": "1.05 MB", "protocol": "tcp", "devs": "0,1,2,3", "threads_per_worker": 1, "times": [ { "wall_clock": 1.4910394318867475, "npartitions": 16 } ], "bandwidths": { "(00,01)": { "25%": "136.34 MB/s", "50%": "156.67 MB/s", "75%": "163.32 MB/s", "total_nbytes": "150.00 MB" } } }