Yeah, I'm sure binding each process to a CPU would be a significant
help. Something I've always wanted to quantify but haven't made time for...
Mark
Luke Lonergan wrote:
> One of our customers noticed that there were a high number of NUMA cache
> misses on a quad core opteron system running Bizgres MPP resulting in about
> a 15% performance hit. We use a process-based parallelization approach and
> we can guess that there's context switching due to the high degree of
> pipeline parallelism in our executions plans. Each context switch likely
> switches a process away from the CPU with local memory, resulting in the
> NUMA cache misses.
>
> The answer for us is to bind each process to a CPU. Might that help in
> running DBT-2?
>
> - Luke
>
>
> On 10/10/06 9:40 AM, "Mark Wong" <markw@osdl.org> wrote:
>
>> Luke Lonergan wrote:
>>> +1
>>>
>>> Mark, can you quantify the impact of not running with IRQ balancing enabled?
>> Whoops, look like performance was due more to enabling the
>> --enable-thread-safe flag.
>>
>> IRQ balancing on : 7086.75
>> http://dbt.osdl.org/dbt/dbt2dev/results/dev4-015/158/
>> IRQ balancing off: 7057.90
>> http://dbt.osdl.org/dbt/dbt2dev/results/dev4-015/163/
>>
>> The interrupt charts look completely different. There's too much stuff
>> on the chart to determine what interrupts are from what though. :( It
>> needs to be redone per processor (as opposed to per interrupt per
>> processor) to be more useful in determining if one processor is
>> overloaded due to interrupts.
>>
>> http://dbt.osdl.org/dbt/dbt2dev/results/dev4-015/158/report/sar/sar-intr.png
>> http://dbt.osdl.org/dbt/dbt2dev/results/dev4-015/163/report/sar/sar-intr.png
>>
>> But the sum of all the interrupts handled are close between tests so it
>> seems clear no single processor was overloaded:
>>
>> http://dbt.osdl.org/dbt/dbt2dev/results/dev4-015/158/report/sar/sar-intr_s.png
>> http://dbt.osdl.org/dbt/dbt2dev/results/dev4-015/163/report/sar/sar-intr_s.png
>>
>> Mark
>>
>
>
