[r-t] Composition search?
alan.burlison at gmail.com
Wed Nov 25 10:15:20 UTC 2015
On 25/11/2015 00:49, Alex Dicks wrote:
> How much speed up do you get with that level of SMT? I've never been
> convinced that Intel's hyperthreading makes that much difference, and have
> benchmarked C compilation on a 10-core/20-thread build server and found a
> plateau at "make -j10". I think it helps a bit if you miss cache often, or
> tend to use machine different instructions across threads/processes. But
> maybe SPARC has more SMT benefits.
It all depends. It's safe to say it's never linear and it's never easy
to achieve. And cache misses are almost never good.
Generally what happens is you end up chasing the bottleneck from one
place to another - for example in your compilation scenario it may well
have been the disk that became a bottleneck, and not disk bandwidth
which is what people normally look at, but IOs/second.
More information about the ringing-theory