I recently had a run-in with a 3x performance diff for @rustlang on the new @amd Threadripper with multi threadded applications. It seems to be related to the layout, I wonder if anyone is interested in me digging deeper into this?
-
-
3x performance improvement on Threadripper?

1 reply 0 retweets 1 like -
Not over intel, over not using explicit core placement. The interesting part is that when left to the linux scheduler performance is horrendously slow.
1 reply 0 retweets 0 likes -
Does it look like Linux is not taking advantage of the extra cores? Or is the Rust application itself not taking advantage of the extra cores?
1 reply 0 retweets 0 likes -
Neither, it seems to be inter thread communication cost between CCX’s or CCD’s but I don’t want to assume too much as I’m not 100% sure what the cause is - just that if it’s talking between core 0,1,3 (or 2,3,4 on a 3960X) it’s fast, if you get out of the grouping it’s slow.
2 replies 0 retweets 0 likes -
Replying to @heinz_gies @raymondsiu and
Is linux’s NUMA core model correct? Would expect bad IPC inter-cluster but would also expect Linux to try and keep related threads on the same cluster...
1 reply 0 retweets 2 likes -
Replying to @experquisite @raymondsiu and
Idk yet I think it does at least based on the default conf. The 3900 threadripper expose a single numa node by default afaik so Linux can’t learn much about it.
1 reply 0 retweets 0 likes
Some MPI / HT Condor folk had to perform some tweaks to optimize affinity on threadripper - some of their tests might be worth replicating https://blog.michael.kuron-germany.de/2018/09/amd-ryzen-threadripper-numa-architecture-cpu-affinity-and-htcondor/ …
-
-
Replying to @darachennis @heinz_gies and
the fact that it lies about not being NUMA sounds megabad. that totally shreds cache coherency performance.
1 reply 0 retweets 2 likes -
- 2 more replies
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.