Re: Parallel pg_dump for 9.1

Jeff wrote:
>
> On Mar 30, 2010, at 8:15 AM, Stefan Kaltenbrunner wrote:
>
>> Peter Eisentraut wrote:
>>> On tis, 2010-03-30 at 08:39 +0200, Stefan Kaltenbrunner wrote:
>>>> on fast systems pg_dump is completely CPU bottlenecked
>>> Might be useful to profile why that is. I don't think pg_dump has
>>> historically been developed with CPU efficiency in mind.
>>
>> It's not pg_dump that is the problem - it is COPY that is the limit.
>> In my specific case als the fact that a lot of the columns are bytea
>> adds to the horrible CPU overhead (fixed in 9.0). Still our bulk load
>> & unload performance is still way slower on a per core comparision
>> than a lot of other databases :(
>>
>
> Don't forget the zlib compression used in -Fc (unless you use -Z0) takes
> a fair amount of cpu too.
> I did some tests and it turned out that -Z0 actually took longer than
> -Z1 simply because there was a lot more data to write out, thus I became
> IO bound not CPU bound.
>
> There's a thing called pigz around that is a parallel gzip
> implementation - wonder how much of that could be adapted to pg_dumps
> use as compression does use a considerable amount of time (even at
> -Z1). The biggest problem I can immediately see is that it uses threads.

all my numbers are with -Z0 and it is the backend (COPY and/or index
creation) that is the limit. If you start using compression you are
shifting the load to pg_dump.

Stefan