On Mon, 2007-03-05 at 10:46 -0800, Josh Berkus wrote:
> Tom,
>
> > I seem to recall that we've previously discussed the idea of letting the
> > clock sweep decrement the usage_count before testing for 0, so that a
> > buffer could be reused on the first sweep after it was initially used,
> > but that we rejected it as being a bad idea. But at least with large
> > shared_buffers it doesn't sound like such a bad idea.
> Note, though, that the current algorithm is working very, very well for OLTP
> benchmarks, so we'd want to be careful not to gain performance in one area at
> the expense of another.
Agreed.
What we should also add to the analysis is that this effect only occurs
when only uniform workloads is present, like SeqScan, VACUUM or COPY.
When you have lots of indexed access the scan workloads don't have as
much effect on the cache pollution as we are seeing in these tests.
Itakgaki-san and I were discussing in January the idea of cache-looping,
whereby a process begins to reuse its own buffers in a ring of ~32
buffers. When we cycle back round, if usage_count==1 then we assume that
we can reuse that buffer. This avoids cache swamping for read and write
workloads, plus avoids too-frequent WAL writing for VACUUM.
It would be simple to implement the ring buffer and enable/disable it
with a hint StrategyHintCyclicBufferReuse() in a similar manner to the
hint VACUUM provides now.
This would maintain the beneficial behaviour for OLTP, while keeping
data within the L2 cache for DSS and bulk workloads.
-- Simon Riggs EnterpriseDB http://www.enterprisedb.com