Re: Poor thread performance on Linux vs. Solaris

Ingo Molnar <[email protected]> Mon, 8 Sep 2003 23:23:31 +0200 (CEST)

Newsgroups	gmane.comp.lib.phil
Message-ID	<[email protected]>

On Mon, 8 Sep 2003, Bill Soudan wrote:

> c0147b60 3540     2.84413     poison_obj
> c0147bf0 4628     3.71825     scan_poisoned_obj

btw., you seem to have CONFIG_DEBUG_SLAB enabled, right? I'd disable it.

> c011e270 4046     3.25066     kunmap_atomic
> c011e1e0 4955     3.98097     kmap_atomic

hm, CONFIG_HIGHPTE enabled?

> c0138100 24115    19.3746     do_futex
> c0138bcd 46827    37.622      .text.lock.futex

note that a higher .text.lock.futex count might be natural if multiple
threads use the _very same_ futex. If there's high contention on that lock
then it will inevitably cause that contention to show up somewhere - on
NPTL it will show up as a higher context-switch count and higher futex
spinlock contention.

btw., what is the total count? It's needed to see the relative overhead of
futex locking.

	Ingo