Обсуждение: Libpq optimization

Поиск
Список
Период
Сортировка

Libpq optimization

От
"Alon Goldshuv"
Дата:
In the libpq COPY interface function PQputCopyData():

/** Check for NOTICE messages coming back from the server.  Since the* server might generate multiple notices during
theCOPY, we have to* consume those in a reasonably prompt fashion to prevent the comm* buffers from filling up and
possiblyblocking the server.*/
 
if (!PQconsumeInput(conn))   return -1;                /* I/O failure */
parseInput(conn);


I moved it to a different location, just a bit further, after the check for
"is output buffer full and we are ready to flush?" in the same function
       if ((conn->outBufSize - conn->outCount - 5) < nbytes)       {           <here>           ....           ....
 }
 


As the code comment suggests, it is extremely important to consume incoming
messages from the server to prevent deadlock. However we should only worry
about it before sending data out. Most calls to PQputCopyData don't actually
send any data but just place it in the out buffer and return. Therefore we
can perform the consumeinput/parseinput right before flushing, instead of
reading from the server every time we call PQputCopyData and not send
anything (which happens probably in 99% of the time).

Right? Or am I missing something.

This change improves COPY performance.

thx
Alon.




Re: Libpq optimization

От
Tom Lane
Дата:
"Alon Goldshuv" <agoldshuv@greenplum.com> writes:
> As the code comment suggests, it is extremely important to consume incoming
> messages from the server to prevent deadlock. However we should only worry
> about it before sending data out.

And, unfortunately, you've broken it.  The pqFlush call visible in that
routine is not the only place that may try to send data (see also
pqPutMsgEnd).
        regards, tom lane


Re: Libpq optimization

От
"Alon Goldshuv"
Дата:
Tom,

> And, unfortunately, you've broken it.  The pqFlush call visible in that
> routine is not the only place that may try to send data (see also
> pqPutMsgEnd).

You are right, thanks for pointing that out.

Still, in pqPutMsgEnd we will be sending data only after 8K is reached,
which is about once in 80 for a 100 byte row size...
Alon.