Robert Haas <robertmhaas@gmail.com> wrote:
> 1. All of the SSI patches have been dealt with.
I'll add the non-serializable UPDATE performance issue. Dan has
been benchmarking to try to find a worst case; I don't want to speak
for him too much, but as he was headed off to lecture a class he
sent me results so far, and with beta so close I figure I should
pass along a rough outline. The worst case he has been able to
construct so far was running 32 active processes on a 16 processor
machine in an update-mostly mix against a database on tmpfs (so no
disk writes) on a dataset which fits inside shared_memory. This was
able to generate enough contention on an exclusive LW lock to cause
a 0.7% slowdown.
Speaking for myself, I believe we'll be able to provide a very small
patch to eliminate this. Probably today or tomorrow. While in a
less extreme runtime environment it would probably be hard to pick
out a performance impact in the normal noise, I expect the fix to be
small and safe enough to be worth doing.
I do feel that it would be good to apply the one-line fix Heikki
posted, which is orthogonal and needed in any event. That would
give a little time for others to easily test it before beta.
-Kevin