Thread: Getting "ERROR: no unpinned buffers available" on HEAD, should I investigate?

Hi

When I run my FK concurrency test suite on an unpatched HEAD, I'm getting quite a lot of "no unpinned buffers
available"errors. 

Increasing shared_buffers from 32MB (the default) to 64MB makes the errors go away, as does setting fsync=off.

I'm not sure how many buffers a particular backend can hold at any time. I'd have though it'd be something like one per
relationand scan-in-progress, with indices counting as relations. In that case, with max_connections=100, my tests
shouldn'tpin more than a couple of hundred buffers at a time. That, however, amounts to about 1MB or so of pinned
buffers,so it shouldn't break with shared_buffers=32MB. 

I'm also confused by fsync=off making a difference. That should make writing dirty buffers slower, but would it affect
thenumber of buffers pinned? 

In short, I'm wondering whether I might have hit a bug, or if I should just increase shared_buffers and move on.

best regards,
Florian Pflug

Excerpts from Florian Pflug's message of mar dic 14 21:51:47 -0300 2010:
> Hi
> 
> When I run my FK concurrency test suite on an unpatched HEAD, I'm getting quite a lot of "no unpinned buffers
available"errors.
 
> 
> Increasing shared_buffers from 32MB (the default) to 64MB makes the errors go away, as does setting fsync=off.
> 
> I'm not sure how many buffers a particular backend can hold at any time. I'd have though it'd be something like one
perrelation and scan-in-progress, with indices counting as relations. In that case, with max_connections=100, my tests
shouldn'tpin more than a couple of hundred buffers at a time. That, however, amounts to about 1MB or so of pinned
buffers,so it shouldn't break with shared_buffers=32MB.
 
> 
> I'm also confused by fsync=off making a difference. That should make writing dirty buffers slower, but would it
affectthe number of buffers pinned?
 
> 
> In short, I'm wondering whether I might have hit a bug, or if I should just increase shared_buffers and move on.

I remember a report of this problem some time ago, and I wasn't able to
explain it -- my calculations said, like yours, that the maximum amount
of pinned buffers I expected shouldn't have been enough to consume the
whole pool.

It would be good to trace down to the root cause.

-- 
Álvaro Herrera <alvherre@commandprompt.com>
The PostgreSQL Company - Command Prompt, Inc.
PostgreSQL Replication, Consulting, Custom Development, 24x7 support