On Wed, Oct 20, 2010 at 1:13 PM, Robert Haas <robertmhaas@gmail.com> wrote:
> On Wed, Oct 20, 2010 at 3:47 PM, daveg <daveg@sonic.net> wrote:
>> On Wed, Oct 20, 2010 at 12:28:25PM -0700, Greg Stark wrote:
>>> On Wed, Oct 20, 2010 at 12:17 PM, Greg Stark <gsstark@mit.edu> wrote:
>>> > I don't think it's a big cost once all the processes
>>> > have been forked if you're reusing them beyond perhaps slightly more
>>> > efficient cache usage.
>>>
>>> Hm, this site claims to get a 13% win just from the reduced tlb misses
>>> using a preload hack with Pg 8.2. That would be pretty substantial.
>>>
>>> http://oss.linbit.com/hugetlb/
>>
>> That was my motivation in trying a patch. TLB misses can be a substantial
>> overhead. I'm not current on the state of play, but working at Sun's
>> benchmark lab on a DB TPC-B benchmark something for the first generation
>> of MP systems, something like 30% of all bus traffic was TLB misses. The
>> next iteration of the hardward had a much larger TLB.
>>
>> I have a client with 512GB memory systems, currently with 128GB configured
>> as postgresql buffer cache. Which is 32M TLB entires trying to fit in the
>> few dozed cpu TLB slots. I suspect there may be some contention.
>>
>> I'll benchmark of course.
>
> Do you mean 128GB shared buffers, or shared buffers + OS cache? I
> think that the general wisdom is that performance tails off beyond
> 8-10GB of shared buffers anyway, so a performance improvement on 128GB
> shared buffers might not mean much unless you can also show that 128GB
> shared buffers actually performs better than some smaller amount.
I'm sure someone will correct me if I'm wrong, but when I looked at
this a couple years ago I believe a side effect of using hugetlbs is
that these segments are never swapped out.
I made a weak attempt to patch postgres to use hugetlbs when
allocating shared memory. If I can find that patch I'll send it out..
Mark