On Fri, Nov 13, 2009 at 11:08 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
(By the way, has anyone yet tried to
> compare the speed of this implementation to the old code?)
I quickly hacked pgbench to take a custom script on connection (for
listen), and make pgbench'd 'notify x'; with all clients doing 'listen
x'.
The old method (measured on a 4 core high performance server) has
severe scaling issues due to table bloat (we knew that):
./pgbench -c 10 -t 1000 -n -b listen.sql -f notify.sql
run #1 tps = 1364.948079 (including connections establishing)
run #2 tps = 573.988495 (including connections establishing)
<vac full pg_listener>
./pgbench -c 50 -t 200 -n -b listen.sql -f notify.sql
tps = 844.033498 (including connections establishing)
new method on my dual core workstation (max payload 128):
./pgbench -c 10 -t 10000 -n -b listen.sql -f notify.sql -hlocalhost postgres
tps = 16343.012373 (including connections establishing)
./pgbench -c 20 -t 5000 -n -b listen.sql -f notify.sql -hlocalhost postgres
tps = 7642.104941 (including connections establishing)
./pgbench -c 50 -t 5000 -n -b listen.sql -f notify.sql -hlocalhost postgres
tps = 3184.049268 (including connections establishing)
max payload 2048:
./pgbench -c 10 -t 10000 -n -b listen.sql -f notify.sql -hlocalhost postgres
tps = 12062.906610 (including connections establishing)
./pgbench -c 20 -t 5000 -n -b listen.sql -f notify.sql -hlocalhost postgres
tps = 7229.505869 (including connections establishing)
./pgbench -c 50 -t 5000 -n -b listen.sql -f notify.sql -hlocalhost postgres
tps = 3219.511372 (including connections establishing)
getting sporadic 'LOG: could not send data to client: Broken pipe'
throughout the test.
merlin
