RE: Poor thread performance on Linux vs. Solaris

On Tue, 9 Sep 2003, Hu, Boris wrote:

> Try the futex_q_lock-0.2 patch. It is also against linux-2.6.0-test4.

indeed it would be nice to try this patch.

in addition, the 'freeze' could be something unrelated, which should be
profiled via the snapshot method i suggested. Since there's almost no
scheduling during such 'freeze' periods, it's unlikely to be a futex
operation (which likely results in a context-switch). Or if it's a futex
operation then it's some sort of bug.

anyway, i do think it's two separate issues - on one side the regular
futex slowpath overhead with your 10-15k/sec context-switches (ie. a lock
contention every 70 usecs or so), on the other side the 'freeze'
phenomenon. The futex locking overhead is likely caused by the non-freeze
activities (freezes are fairly short compared to the total test).

	Ingo