On Tue, 2007-03-06 at 22:32 -0500, Luke Lonergan wrote:
> Incidentally, we tried triggering NTA (L2 cache bypass)
> unconditionally and in various patterns and did not see the
> substantial gain as with reducing the working set size.
>
> My conclusion: Fixing the OS is not sufficient to alleviate the issue.
> We see a 2x penalty (1700MB/s versus 3500MB/s) at the higher data
> rates due to this effect.
>
I've implemented buffer recycling, as previously described, patch being
posted now to -patches as "scan_recycle_buffers".
This version includes buffer recycling
- for SeqScans larger than shared buffers, with the objective of
improving L2 cache efficiency *and* reducing the effects of shared
buffer cache spoiling (both as previously discussed on this thread)
- for VACUUMs of any size, with the objective of reducing WAL thrashing
whilst keeping VACUUM's behaviour of not spoiling the buffer cache (as
originally suggested by Itagaki-san, just with a different
implementation).
Behaviour is not activated by default in this patch. To request buffer
recycling, set the USERSET GUCSET scan_recycle_buffers = N
tested with 1,4,8,16, but only > 8 seems sensible, IMHO.
Patch effects StrategyGetBuffer, so only effects the disk->cache path.
The idea is that if its already in shared buffer cache then we get
substantial benefit already and nothing else is needed. So for the
general case, the patch adds a single if test into the I/O path.
The parameter is picked up at the start of SeqScan and VACUUM
(currently). Any change mid-scan will be ignored.
IMHO its possible to do this and to allow Synch Scans at the same time,
with some thought. There is no need for us to rely on cache spoiling
behaviour of scans to implement that feature as well.
Independent performance tests requested, so that we can discuss this
objectively.
-- Simon Riggs EnterpriseDB http://www.enterprisedb.com
