hkaiser changed the topic of #ste||ar to: STE||AR: Systems Technology, Emergent Parallelism, and Algorithm Research | stellar-group.org | HPX: A cure for performance impaired parallel applications | github.com/STEllAR-GROUP/hpx | This channel is logged: irclog.cct.lsu.edu
<hkaiser> PatrickDiehl[m]: +1
diehlpk has joined #ste||ar
K-ballo has quit [Quit: K-ballo]
hkaiser has quit [Quit: Bye!]
diehlpk has quit [Quit: Leaving.]
K-ballo has joined #ste||ar
hkaiser has joined #ste||ar
<jedi18[m]> hkaiser: What chunk sizes do you recommend I test it with? I tried size/48 and size/96 size chunks but those perform the same/worse than default
<hkaiser> jedi18[m]: I think reducing the number of chunks might help
<hkaiser> i.e. one chunk per core
<hkaiser> especially for small sequence sizes
<jedi18[m]> Ok so since there are 48 cores, size/48 should do that right?
<jedi18[m]> https://github.com/Jedi18/scan_benchmarks/tree/main/varying_chunk_size default still seems to perform better
<hkaiser> jedi18[m]: ok
<hkaiser> jedi18[m]: can you try running with --hpx:threads=24 or so?
<hkaiser> btw, the default is cores*4 chunks
<hkaiser> also, could you add the L1/L2/L3 cache sizes as vertical lines on the graph? that might help understanding the drops in scaling
<jedi18[m]> Oh ok sure
Yorlik has joined #ste||ar
<gnikunj[m]> hkaiser yes, I'm glad my university recognized our work! I'll get down to Louisiana this December. So let me treat you to a beer ;)
<hkaiser> gnikunj[m]: +1
tufei has joined #ste||ar