On Wed, Dec 21, 2011 at 5:33 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> ... while the main buffer manager is
>> content with some loosey-goosey approximation of recency, the SLRU
>> code makes a fervent attempt at strict LRU (slightly compromised for
>> the sake of reduced locking in SimpleLruReadPage_Readonly).
>
> Oh btw, I haven't looked at that code recently, but I have a nasty
> feeling that there are parts of it that assume that the number of
> buffers it is managing is fairly small. Cranking up the number
> might require more work than just changing the value.
My memory was that you'd said benchmarks showed NUM_CLOG_BUFFERS needs
to be low enough to allow fast lookups, since the lookups don't use an
LRU they just scan all buffers. Indeed, it was your objection that
stopped NUM_CLOG_BUFFERS being increased many years before this.
With the increased performance we have now, I don't think increasing
that alone will be that useful since it doesn't solve all of the
problems and (I am told) likely increases lookup speed.
The full list of clog problems I'm aware of is: raw lookup speed,
multi-user contention, writes at checkpoint and new xid allocation.
Would it be better just to have multiple SLRUs dedicated to the clog?
Simply partition things so we have 2^N sets of everything, and we look
up the xid in partition (xid % (2^N)). That would overcome all of the
problems, not just lookup, in exactly the same way that we partitioned
the buffer and lock manager. We would use a graduated offset on the
page to avoid zeroing pages at the same time. Clog size wouldn't
increase, we'd have the same number of bits, just spread across 2^N
files. We'd have more pages too, but that's not a bad thing since it
spreads out the contention.
Code-wise, those changes would be isolated to clog.c only, probably a
days work if you like the idea.
--
Simon Riggs http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services