Stefan Kaltenbrunner wrote:
> Tom Lane wrote:
>> "Joshua D. Drake" <jd@commandprompt.com> writes:
>>> Yeah, that... is beyond my abilities. Well reading it is anyway. I can
>>> provide any information people want though. Tom? Greg? Andrew? Somebody?
>>> What information do you want from me to help you track this down?
>>
>> The vmstat output you showed before said user CPU time was only around
>> 12%, which seems to destroy Luke's thesis that data conversion time
>> is the problem. IIRC we were speculating that data was being written
>> in a pattern that required a lot of seeking thus ruining throughput,
>> but we didn't have any hard evidence of that. Did you do the strace'ing
>> I suggested?
>
>
> Joshua said this is an 8 core box - so 12% would about equal to one core
> being maxed out completely CPU wise which seems to correlate fairly well
> with my experience (COPY being CPU bottlenecked on fast hardware without
> any real culprit to find in profiles).
Right. I mentioned previously that I was able to gain what could be
considered barely acceptable performance break breaking up a custom dump
into multiple pieces and restoring across multiple connections.
I have Alvaro reviewing the strace. I can't post it without some cleansing.
Sincerely,
Joshua D. Drake