Closed npadmana closed 4 years ago
Feel free to assign this to me.
I see the same performance differences for the UPC and MPI versions too, so I would chalk this up to machine differences. We were expecting Swan to be faster because it's 44-core 2.2 GHz Broadwell chips vs 36-core 2.1 GHz Broadwell chips on Crystal. I don't understand the difference yet, but it's consistent across reference versions, so I don't think this is anything to be concerned about for Chapel.
Below are Chapel/UPC times for 32 nodes for size D/E:
Config | Swan | Crystal |
---|---|---|
Chapel | 11.61 s | 10.39 s |
UPC | 13.79 s | 11.97 s |
Config | Swan | Crystal |
---|---|---|
Chapel | 98.68 s | 88.04 s |
UPC | 115.67 s | 104.71 s |
I think we can close this.
@npadmana has been seeing noticeably slower performance on swan than the benchmarks @ronawho sees on crystal.
Not urgent, but would be good to track down the differences.