Re: Slow network retrieves

<wespvp(at)syntegra(dot)com> writes:
> On 5/9/04 9:32 AM, "Tom Lane" <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> Are you sure it is a network problem?

> Yes, it is definitely due to the network latency even though that latency is
> very small. Here it is running locally:
> [ about 20000 records/sec ]

Okay, I just wanted to verify that we weren't overlooking any other
sorts of bottleneck. But the numbers you quote make sense as a network
issue: 33 seconds for 10000 records is 3.03 msec per record, and since
you say the measured ping time is 3 msec, it appears that FETCH has
just about the same response time as a ping ;-). So you can't really
complain about it. The only way to do better will be to batch multiple
fetches into one network round trip.

> A Pro*C program I recently ported from Oracle to PostgreSQL showed this
> difference. In Pro*C you can load an array with rows to insert, then issue
> a single INSERT request passing it the array. As far as I can tell, in
> PostgreSQL ecpg (or other) you have to execute one request per record.

The usual way to batch multiple insertions is with COPY IN. The usual
way to batch a fetch is just to SELECT the whole thing; or if that is
too much data to snarf at once, use a cursor with "FETCH n" requests.
I am not sure how either of these techniques map into ecpg though.
If you want to use ecpg then I'd suggest bringing up the question on
pgsql-interfaces --- the ecpg gurus are more likely to be paying
attention over there.

> ... It appears that COPY works like this, but you can't
> control what is returned and you have to know the column order.

True, COPY OUT is only designed to return all the rows of a table.
However, in recent versions you can specify what columns you want
in a COPY. It's still no substitute for SELECT...

regards, tom lane