Fix the difference in 2D partitioning of GPUs in python and C++ #1950
Conversation
seunghwak
added
3 - Ready for Review
improvement
…ython_2d_partition
|
rerun tests Rerunning to get latest updates that should fix build problem. |
Codecov Report
@@ Coverage Diff @@
## branch-22.02 #1950 +/- ##
===============================================
Coverage ? 70.28%
===============================================
Files ? 143
Lines ? 8922
Branches ? 0
===============================================
Hits ? 6271
Misses ? 2651
Partials ? 0 Continue to review full report at Codecov.
|
| pcols = pcols - 1 | ||
| return int(ngpus/pcols), pcols | ||
| prows = int(math.sqrt(ngpus)) | ||
| while ngpus % prows != 0: |
There was a problem hiding this comment.
this will be a very fast loop, but you should have just been able to subtract the delta of the modulo result from prows rather than looping.
There was a problem hiding this comment.
Can we do this? I am not sure I am fully understanding your intention, but here prows is varying.
Say ngpus = 14. Then, int(sqrt(14)) will be 3. 14 % 3 will be 2. prows (3) - 2 will be 1. But we want prows to be 2.
This code is based on the discussions from https://stackoverflow.com/questions/16266931/input-an-integer-find-the-two-closest-integers-which-when-multiplied-equal-th
|
@gpucibot merge |
If sqrt(# GPUs) is not an integer, prows > pcols in python and prows < pcols in C++.
The C++'s choice leads to better performance (especially for BFS & SSSP and on the systems with low inter-GPU communication bandwidth) and this reduces the difference between python and C++ testing/benchmarking.