Performance!

[Kris Kennaway]

Krassimir Slavchev wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> 
> Kris Kennaway wrote:
>> Krassimir Slavchev wrote:
>>> -----BEGIN PGP SIGNED MESSAGE-----
>>> Hash: SHA1
>>>
>>> Hello,
>>>
>>> Kris Kennaway wrote:
>>>> Krassimir Slavchev wrote:
>>>>> -----BEGIN PGP SIGNED MESSAGE-----
>>>>> Hash: SHA1
>>>>>
>>>>> Hello,
>>>>>
>>>>> Here are lock profiling results with select patch applied.
>>>> OK, you are doing I/O over TCP.  Are you sure you are using TCP on both
>>>> systems?  Linux may not be defaulting to TCP transport for local
>>>> queries.
>>>>
>>>> Add --pgsql-host="" to your sysbench command line to make it communicate
>>>> over a local domain socket, which is much more efficient.
>>>>
>>>> Kris
>>>>
>>> Hmm, Yes linux uses local domain sockets!
>>> Here are results using local domain sockets on FreeBSD too:
>>> #threads        #tranzactions/sec
>>> 1               728
>>> 5               2996
>>> 10              5301
>>> 20              3931
>>> 40              2466
>>> 60              1852
>>> 80              1424
>>> 100             1216
>>>
>>> Just to remember:
>>> Linux (2.6.18)
>>> #threads        #transactions/sec
>>> 1               693
>>> 5               3539
>>> 10              5789
>>> 20              5791
>>> 40              5661
>>> 60              5517
>>> 80              5401
>>> 100             5319
>>>
>>> I have results using Fedora 8 on the same hardware:
>>> Linux (2.6.23)
>>> #threads        #transactions/sec
>>> 1               740
>>> 5               2675
>>> 10              6486
>>> 20              6893
>>> 40              6623
>>> 60              6623
>>> 80              6522
>>> 100             6417
>>>
>>> If we look at the results with up to 10 threads the performance of
>>> FreeBSD is very good.
>>> May be something can be tuned for number of threads > number of CPUs?
>>>
>>> Are you interested in lock profiling statistics with more threads than
>>> the number of CPUs?
>> Yes, it's still performing anomalously.  Glad we're making progress
>> though :)
>>
>> Kris
>>
> 
> Okay, but how many threads will be more useful to test with?

Let's start with 8 and say 40.  The two problems seem to be slightly 
lower peak and poor scaling above peak.  They might be the same, or they 
might be different.

kris