Low throughput about costa HOT 5 CLOSED

update: costa-2.1 gives the same low performance as costa-2.0

from costa.

Hi @toxa81

After a lot of benchmarking, I finally realized what was causing the low throughput in these cases:

1. Loop unrolling makes the copy/transpose kernels 2x faster.
2. After loop unrolling, 70% of time is spent in temporary buffer allocation. These buffers are used for packing/unpacking the messages and are allocated/deallocated each time a transform() is invoked. We can fix this by using a memory pool.

So, the good news is that this should not be too hard to fix - will keep you updated.

from costa.

Hi Anton,

The throughput should now be faster - it goes up to 7.3GB/s. The test file you sent me is in examples/copy_bench.cpp:

OMP_NUM_THREADS=12 srun -N 2 -n 2 ./examples/copy_bench
[costa::transform] throughput: 3.01352 Gb/sec <- loop unrolling boosts the perf. 2x
[costa::transform] throughput: 5.7582 Gb/sec  <- packing/unpacking buffers reused
[costa::transform] throughput: 4.71008 Gb/sec
[costa::transform] throughput: 7.27846 Gb/sec
[costa::transform] throughput: 4.71543 Gb/sec
[costa::transform] throughput: 7.26714 Gb/sec
[costa::transform] throughput: 5.76084 Gb/sec
[costa::transform] throughput: 7.27839 Gb/sec
[costa::transform] throughput: 7.27399 Gb/sec
[costa::transform] throughput: 4.71403 Gb/sec

from costa.

Hi Marco! I think it looks better now. I will test it in the production runs and let you know.

from costa.

Thanks Anton! I think the limiting factor here is the packing/unpacking which takes up half of the bandwidth, but is still faster than letting MPI pack/unpack it.

from costa.