Re: [PATCH] jbd2: Use atomic variables to avoid taking t_handle_lock in jbd2_journal_stop

[Eric Whitney <eric.whitney@hp.com>]

Ted Ts'o wrote:
> On Mon, Aug 02, 2010 at 07:52:29PM -0700, john stultz wrote:
>> With the non-vfs scalability patched kernels, we see that the j_state
>> lock and atomic changes pull start_this_handle out of the top contender
>> handle, but there is still quite a large amount of contention on the
>> dput paths.
>>
>> So yea, the change does help, but its just not the top cause of
>> contention when aren't using the vfs patches, so we don't see as much
>> benefit at this point.
> 
> Great, thanks for uploading the lockstats.  Since dbench is so
> metadata heavy, it makes a lot of sense that further jbd2
> optimizations probably won't make much difference until the VFS
> bottlenecks can be solved.
> 
> Other benchmarks, such as the FFSB benchmarks used by Steven Pratt and
> Eric Whitney, would probably show more of a difference.
> 
> In any case, I've just sent two more patches which completely remove
> any exclusive spinlocks from start_this_handle() by converting
> j_state_lock to a rwlock_t, and dropping the need to take
> t_handle_lock.  This will add more cache line bouncing, so on NUMA
> workloads this may make things worse, but I guess we'll have to see.
> Anyone have access to an SGI Altix?  I'm assuming the old Sequent NUMA
> boxes are long gone by now...
> 
> 					- Ted

Ted:

The 48 core system I'm running on is an eight node NUMA box with three 
hop worst case latency, and it tends to let you know if you're bouncing 
cache lines too enthusiastically.  If someone's got access to a larger 
system with more hops across the topology, that would naturally be even 
better.

I'm taking my 2.6.35 ext4 baseline now.  It takes me about 36 hours of 
running time to get in a complete set of runs, so with luck I should 
have data and lockstats to post in a few days.

Eric

P.S.  And yes, I think I'll make a set of non-accidental no-journal runs 
this time as well...