On Mon, Aug 10, 2015 at 3:09 PM, Amit Kapila <amit.kapila16@gmail.com> wrote:
> On Thu, Aug 6, 2015 at 12:24 PM, Haribabu Kommi <kommi.haribabu@gmail.com>
> wrote:
>
> What is the configuration for test (RAM of m/c, shared_buffers,
> scale_factor, etc.)?
Here are the details:
CPU - 16 core, RAM - 252 GB
shared_buffers - 1700MB, buffer_cache_ratio - 70
wal_buffers - 16MB, synchronous_commit - off
checkpoint_timeout - 15min, max_wal_size - 5GB.
pgbench scale factor - 75 (1GB)
Load test table size - 1GB
>> Threads Head Patched Diff
>> 1 3123 3238 3.68%
>> 2 5997 6261 4.40%
>> 4 11102 11407 2.75%
>>
>> I am suspecting that, this may because of buffer locks that are
>> causing the problem.
>> where as in older approach of different buffer pools, each buffer pool
>> have it's own locks.
>> I will try to collect the profile output and analyze the same.
>>
>> Any better ideas?
>>
>
> I think you should try to find out during test, for how many many pages,
> it needs to perform clocksweep (add some new counter like
> numBufferBackendClocksweep in BufferStrategyControl to find out the
> same). By theory your patch should reduce the number of times it needs
> to perform clock sweep.
>
> I think in this approach even if you make some buffers as non-replaceable
> (buffers for which BM_BUFFER_CACHE_PAGE is set), still clock sweep
> needs to access all the buffers. I think we might want to find some way to
> reduce that if this idea helps.
>
> Another thing is that, this idea looks somewhat similar (although not same)
> to current Ring Buffer concept, where Buffers for particular types of scan
> uses buffers from Ring. I think it is okay to prototype as you have done
> in patch and we can consider to do something on those lines if at all
> this patch's idea helps.
Thanks for the details. I will try the same.
Regards,
Hari Babu
Fujitsu Australia