I was testing the spinlock patches that Tom Lane posted last month
on a quad opteron system running Suse 9.2 for x86_64. The test sql and
database was from a real application of ours and I was interested in seeing
what effect the patches might have. The database is entirely RAM resident and
nothing else was running while I ran the tests. The results are puzzling (to
me anyway).
There was a big boost in performance when the concurrency level was less than
or equal to 4 and a much smaller boost when the concurrency level was greater
than 3. The puzzling part was the context switches though. Why the big
falloff above 5? In any case these results explain some performance meltdowns
we have been seeing. When the concurrency goes above 5 the performance falls
off so rapidly that we would be better running serially on a single processor
and the system has a hard time recovering.
with patch
N time cs from vmstat
1 37 50-100
2 49 70k-80k
4 79 290k-300k
5 183 6k-10k
8 299 4k-6k
without patch
N time cs from vmstat
1 36 50-100
2 53 55k-60k
4 185 45k-55k
5 218 7k-10k
8 315 4k-6k