Page 1 of 1
Opinions needed on a performance test
Posted: Wed Jul 04, 2012 8:05 am
by IanSeyler
I've been doing some benchmark tests between Linux and BareMetal OS and have found some interesting results. With the PrimeSMP program being tested on a dual core machine I saw the almost 2x performance increase I expected. However, while running on a 16-core system with Linux this increase does not keep up with the theoretical maximum. Is Linux this bad when it comes to handling multiple cores?
Information on the test is here:
http://blog.returninfinity.com
Thanks,
-Ian
Re: Opinions needed on a performance test
Posted: Thu Jul 05, 2012 1:11 am
by Nable
The first thing that i can mention about linux: if you want performance, you have to manually set affinity of your processes. I.e. lock each process (or thread, it's not important here) to it's dedicated core, else scheduler will waste a lot of time migrating processes from core to core. I don't know why he does it, but situation is the same as several years ago: if you start a cpu intensive task, scheduler will move it from core to core instead of keeping it on some fixed one.
Second fact: linux distros have many unneeded stuff out-of-box, so you should stop daemons that you are not critical, it can save you several precents of performance.
Third fact is about preemption and timer interrupts, see your kernel config for obvious details.