Re: [PATCH] hugepage: allow parallelization of the hugepage faultpath

[Davidlohr Bueso]

On Thu, 2013-07-18 at 17:19 -0700, Davidlohr Bueso wrote:
> On Thu, 2013-07-18 at 18:07 +0900, Joonsoo Kim wrote:
> > On Wed, Jul 17, 2013 at 12:50:25PM -0700, Davidlohr Bueso wrote:
> > 
> > > From: Davidlohr Bueso <davidlohr.bueso@xxxxxx>
> > > 
> > > - Cleaned up and forward ported to Linus' latest.
> > > - Cache aligned mutexes.
> > > - Keep non SMP systems using a single mutex.
> > > 
> > > It was found that this mutex can become quite contended
> > > during the early phases of large databases which make use of huge pages - for instance
> > > startup and initial runs. One clear example is a 1.5Gb Oracle database, where lockstat
> > > reports that this mutex can be one of the top 5 most contended locks in the kernel during
> > > the first few minutes:
> > > 
> > >     	     hugetlb_instantiation_mutex:   10678     10678
> > >              ---------------------------
> > >              hugetlb_instantiation_mutex    10678  [<ffffffff8115e14e>] hugetlb_fault+0x9e/0x340
> > >              ---------------------------
> > >              hugetlb_instantiation_mutex    10678  [<ffffffff8115e14e>] hugetlb_fault+0x9e/0x340
> > > 
> > > contentions:          10678
> > > acquisitions:         99476
> > > waittime-total: 76888911.01 us
> > 
> > Hello,
> > I have a question :)
> > 
> > So, each contention takes 7.6 ms in your result.
> 
> Well, that's the total wait time. I can see your concern, but no, things
> aren't *that* bad. The average amount of time spent waiting for the lock
> would be 76888911.01/10678 = 7200us
> 

Long day, I think this is what you were saying for the start :)

> > Do you map this area with VM_NORESERVE?
> > If we map with VM_RESERVE, when page fault, we just dequeue a huge page from a queue and clear
> > a page and then map it to a page table. So I guess, it shouldn't take so long.
> > I'm wondering why it takes so long.
> > 
> 
> I cannot really say. This is proprietary software. AFAICT if Oracle is
> anything like Posgres, than probably no.
> 
> 
> > And do you use 16KB-size hugepage?
> 
> No, 2Mb pages.
> 
> > If so, region handling could takes some times. If you access the area as random order,
> > the number of region can be more than 90000. I guess, this can be one reason to too long
> > waittime.
> > 
> > Thanks.
> 
> Thanks,
> Davidlohr
> 


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/