On Sat, Jan 24, 2015 at 12:24 AM, Joshua D. Drake <
jd@commandprompt.com> wrote:
>
>
> On 01/23/2015 10:44 AM, Jim Nasby wrote:
>>
>> number of workers especially at slightly higher worker count.
>>
>> Those fixed chunk numbers look pretty screwy. 2, 4 and 8 workers make no
>> difference, then suddenly 16 cuts times by 1/2 to 1/3? Then 32 cuts time
>> by another 1/2 to 1/3?
>
There is variation in tests at different worker count but there is
definitely improvement from 0 to 2 worker count (if you refer my
initial mail on this data, with 2 workers there is a benefit of ~20%)
and I think we run the tests in a similar way (like compare 0 and 2
or 0 or 4 or 0 and 8), then the other effects could be minimised and
we might see better consistency, however the general trend with
fixed-chunk seems to be that scanning that way is better.
I think the real benefit with the current approach/patch can be seen
with qualifications (especially costly expression evaluation).
Further, if we want to just get the benefit of parallel I/O, then
I think we can get that by parallelising partition scan where different
table partitions reside on different disk partitions, however that is
a matter of separate patch.
>
> cached? First couple of runs gets the relations into memory?
>
Not entirely, as the table size is double than RAM, so each run
