Subject: yamt-pdpolicy benchmark
To: None <firstname.lastname@example.org>
From: YAMAMOTO Takashi <email@example.com>
Date: 03/14/2006 00:40:20
i've done a simple benchmark for yamt-pdpolicy branch and put the result at:
(i chose postgres because LIRS wins for the workload,
according to the paper. :-)
for y in 71 72 73; do
for x in $(seq 1 30); do
pgbench -S -c $x -t 100 bench
done > $y.txt
x-axis is $x-1.
y-axis is "tps = xxx (excluding connections establishing)".
there are three kernels, "clock", "clock.type", "cp".
all of them are built from the same sources.
- "clock" is a kernel without PDPOLICY_CLOCKPRO,
with the following sysctl settings.
- "clock.type" is the same kernel as "clock",
with default sysctl settings.
- "cp" is a kernel with PDPOLICY_CLOCKPRO.
the machine is XENU with 40MB domain memory.
postgres is "postgresql80-8.0.5" from pkgsrc, with no tweaks.
it's interesting that the first run of "cp" is slow.
maybe it's because it takes some time to identify hot pages,
but i'm not sure.