Closed npadmana closed 4 years ago
I believe these are actually the same (the issue was my confusion on what the UPC benchmarks were called), but it would be worth a quick second confirmation. We can then close this....
@ronawho -- do you mind checking that we're computing the same way?
These look the same to me, but as an extra sanity check I'd take the 3 versions (Chapel, UPC, and MPI) and hardcode the total time to a few different values and make sure the MFlop/s is the same.
Based on the runs in https://github.com/npadmana/DistributedFFT/pull/57 and some sanity checks I did with manually setting the total time, I'm pretty confident that we're reporting the same thing as UPC/MPI references.
I think this can be closed.
Agreed.
33 shows that the UPC benchmark runs faster than our code, however, the MFlops reported is lower. We should make sure that all the codes compute MFlops correctly.
The UPC code is
Here is the Chapel version