Re: [PATCH 0/3 v3] dcache: make it more scalable on large system

[Simo Sorce <simo@redhat.com>]

On Wed, 2013-05-29 at 18:56 +0200, Andi Kleen wrote:
> On Wed, May 29, 2013 at 12:18:09PM -0400, Simo Sorce wrote:
> > On Wed, 2013-05-29 at 11:55 -0400, Waiman Long wrote:
> > 
> > > My patch set consists of 2 different changes. The first one is to avoid 
> > > taking the d_lock lock when updating the reference count in the 
> > > dentries. This particular change also benefit some other workloads that 
> > > are filesystem intensive. One particular example is the short workload 
> > > in the AIM7 benchmark. One of the job type in the short workload is 
> > > "misc_rtns_1" which calls security functions like getpwnam(), 
> > > getpwuid(), getgrgid() a couple of times. These functions open the 
> > > /etc/passwd or /etc/group files, read their content and close the files. 
> > > It is the intensive open/read/close sequence from multiple threads that 
> > > is causing 80%+ contention in the d_lock on a system with large number 
> > > of cores.
> > 
> > To be honest a workload base on /etc/passwd or /etc/group is completely
> > artificial, in actual usage, if you really have  such access you use
> > nscd or sssd with their shared memory caches to completely remove most
> > of the file access.
> 
> I don't fully agree at this point. A lot of things can be tuned away,
> but in practice we want things to perform well out of the box without
> needing all kinds of magic tuning that only 

Phrase seem cut mid-sentence ?

> Also this is just normal file access, nothing special about it.
> It simply has to scale. For all kinds of workloads.
> 
> And it does, just d_path messes it up.

Well there are reasonable workloads and artificial ones, I am just
warning not to use 'that' specific test as a good indicator, if you have
other reasonable workloads that show a similar flow feel free to bring
it up.

> > I have no beef on the rest but repeated access to Nsswitch information
> > is not something you need to optimize at the file system layer and
> > should not be brought up as a point in favor.
> 
> This is about repeated access to arbitrary files.

Ok. I do not want to start a discussion on this, I just pointed out the
specific point was not really a good example hopefully there are others
that justify the patchset.

Simo.

-- 
Simo Sorce * Red Hat, Inc * New York