Re: Write Ahead Logging for Hash Indexes

On Sun, Sep 11, 2016 at 3:01 PM, Mark Kirkwood
<mark(dot)kirkwood(at)catalyst(dot)net(dot)nz> wrote:
> On 11/09/16 19:16, Mark Kirkwood wrote:
>
>>
>>
>> On 11/09/16 17:01, Amit Kapila wrote:
>>>
>>> ...Do you think we can do some read-only
>>> workload benchmarking using this server? If yes, then probably you
>>> can use concurrent hash index patch [1] and cache the metapage patch
>>> [2] (I think Mithun needs to rebase his patch) to do so.
>>>
>>>
>>>
>>> [1] -
>>> https://www.postgresql.org/message-id/CAA4eK1J6b8O4PcEPqRxNYbLVbfToNMJEEm+qn0jZX31-obXrJw@mail.gmail.com
>>> [2] -
>>> https://www.postgresql.org/message-id/CAD__OuhJ29CeBif_fLGe4t9Vj_-cFXBwCXhjO+D_16TXbemY+g@mail.gmail.com
>>>
>>>
>>
>> I can do - are we checking checking for hangs/assertions or comparing
>> patched vs unpatched performance (for the metapage patch)?
>>
>>

The first step is to find any bugs and then do performance testing.
What I wanted for performance testing was to compare HEAD against all
three patches (all three together) of Hash Index (concurrent hash
index, cache the meta page, WAL for hash index). For Head, we want
two set of numbers, one with hash indexes and another with btree
indexes. As Jeff has found few problems, I think it is better to
first fix those before going for performance tests.

>
> So, assuming the latter - testing performance with and without the metapage
> patch:
>
> For my 1st runs:
>
> - cpus 16, ran 16G
> - size 100, clients 32
>
> I'm seeing no difference in performance for read only (-S) pgbench workload
> (with everybody using has indexes). I guess not that surprising as the db
> fites in ram (1.6G and we have 16G). So I'll retry with a bigger dataset
> (suspect size 2000 is needed).
>

I think you should try with -S -M prepared and with various client
counts (8,16,32,64,100...).

--
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com