[Bizgres-general] Statement Queuing take II - Resource Scheduling (Running)

[Mark Kirkwood]

Jim C. Nasby wrote:
> On Mon, Jul 17, 2006 at 11:38:19AM +1200, Mark Kirkwood wrote:
>   
>> There is still several things to look at before any serious testing is 
>> done (e.g. the remaining points in the review), however I was pleased to 
>> notice that I could not detect any difference in performance between 
>> scheduling on or off for a few runs of pgbench.
>>
>> The test used number of clients less than the count limit for queue, so 
>> was only attempting to measure the overhead introduced by the queue lock 
>> and unlock operations, rather than the effect of throttling the number 
>> of executing statements (need the deadlock handling in place before that 
>>     
>
> Do you also have results without your patch in? That's undoubtedly
> something the community will be very interested in.
>   

Good point - I can't recall if I actually tested against unpatched 
Bizgres. A quick retest using pgbench with scale = 10 plus 10 clients on 
Bizgres CVS from 14 June gives:

Bizgres                                        - 328 tps
Bizgres + scheduler patch                      - 340 tps
Bizgres + scheduler patch + scheduling enabled - 320 tps

Repeated runs show similar results, all around 300-350. Clearly pgbench 
is not consistent enough to detect whatever hit the queuing is introducing.

I'm going to spend some time later on getting OSDL dbt2 and dbt3 tests 
going, as these should be able to accurately show us the performance hit 
(dbt2), and possibly demonstrate some of the benefits (dbt3) of this work!

Cheers

Mark