|From:||Alex K <kondratov(dot)aleksey(at)gmail(dot)com>|
|Cc:||Stephen Frost <sfrost(at)snowman(dot)net>, Anastasia Lubennikova <lubennikovaAV(at)gmail(dot)com>, Alexander Korotkov <aekorotkov(at)gmail(dot)com>, David Steele <david(at)pgmasters(dot)net>, Alvaro Herrera <alvherre(at)alvh(dot)no-ip(dot)org>|
|Subject:||Parallel COPY FROM execution|
|Views:||Raw Message | Whole Thread | Download mbox|
I am a GSOC student this year, my initial proposal has been discussed
in the following thread
Patch with COPY FROM errors handling seems to be quite finished, so
I have started thinking about parallelism in COPY FROM, which is the next
point in my proposal.
In order to understand are there any expensive calls in COPY, which
can be executed in parallel, I did a small research. First, please, find
flame graph of the most expensive copy.c calls during the 'COPY FROM file'
attached (copy_from.svg). It reveals, that inevitably serial operations like
CopyReadLine (<15%), heap_multi_insert (~15%) take less than 50% of
time in summary, while remaining operations like heap_form_tuple and
multiple checks inside NextCopyFrom probably can be executed well in parallel.
Second, I have compared an execution time of 'COPY FROM a single large
file (~300 MB, 50000000 lines)' vs. 'COPY FROM four equal parts of the
original file executed in the four parallel processes'. Though it is a
very rough test, it helps to obtain an overall estimation:
Thus, it results in a ~60% performance boost per each x2 multiplication of
parallel processes, which is consistent with the initial estimation.
After several discussions I have two possible solutions on my mind:
1) Simple solution
Let us focus only on the 'COPY FROM file', then it is relatively easy to
implement, just give the same file and offset to each worker.
++ Simple; more reliable solution; probably it will give us the most possible
- - Limited number of use cases. Though 'COPY FROM file' is a frequent case,
even when one use it with psql \copy, client-side file read and stdin
streaming to the backend are actually performed
2) True parallelism
Implement a pool of bg_workers and simple shared_buffer/query. While main
COPY process will read an input data and put raw lines into the query, parallel
bg_workers will take lines from there and process.
++ More general solution; support of various COPY FROM use-cases
- - Much more sophisticated solution; probably less performance boost
compared to 1)
I will be glad to any comments and criticism.
|Next Message||Pavel Stehule||2017-06-30 12:35:46||Re: Parallel COPY FROM execution|
|Previous Message||Jeevan Ladhe||2017-06-30 12:18:53||Re: Adding support for Default partition in partitioning|