Re: New lockless pagecache

[Nick Piggin <nickpiggin@yahoo.com.au>]

Alok kataria wrote:
> Hi Nick,
> 
> I have collected performance numbers for the lock less page cache
> patch on the AIM - IO test.
> The performance numbers are collected for 1-100 tasks 1-50 tasks and
> 90-100 tasks  both for with and without your patch. This was done on
> 2.6.13 kernel.
> There's definite improvement when the tasks are small i.e ~50-70. But
> when the tasks go beyond 80, we see a large performance dip.
> I again profiled the 90-100 runs with spinlock's inlined, but couldn't
> understand the reason behind the performance difference.
> 
> Please find attached the performance numbers as well as the oprofile logs.
> 

Hi Alok,

Thanks very much for doing these numbers. Performance is improved
significantly at smaller numbers of tasks, as you say.

Unfortunately I can't pinpoint the reason why performance drops at
larger numbers. I could assume that the last remaining place that
used read_lock_irq for the tree_lock (wait_on_page_writeback_range)
got hurt when switching to spinlocks, but that would seem vary
unlikely.

I'll have to look into it further.

Thanks,
Nick

-- 
SUSE Labs, Novell Inc.

Send instant messages to your online friends http://au.messenger.yahoo.com