On machines with lots of CPU cores, pgbench can start eating up a lot
of system time. Investigation reveals that the problem is with
random(), which glibc implements like this:
long int
__random ()
{
int32_t retval;
__libc_lock_lock (lock);
(void) __random_r (&unsafe_state, &retval);
__libc_lock_unlock (lock);
return retval;
}
weak_alias (__random, random)
Rather obviously, if you're running enough pgbench threads, you're
going to have a pretty ugly point of contention there. On the 32-core
machine provided by Nate Boley, with my usual 5-minute SELECT-only
test, lazy-vxid and sinval-fastmessages applied, and scale factor 100,
"time" shows that pgbench uses almost as much system time as user
time:
$ time pgbench -n -S -T 300 -c 64 -j 64
transaction type: SELECT only
scaling factor: 100
query mode: simple
number of clients: 64
number of threads: 64
duration: 300 s
number of transactions actually processed: 55319555
tps = 184396.016257 (including connections establishing)
tps = 184410.926840 (excluding connections establishing)
real 5m0.019s
user 21m10.100s
sys 17m45.480s
I patched it to use random_r() - the patch is attached - and here are
the (rather gratifying) results of that test:
$ time ./pgbench -n -S -T 300 -c 64 -j 64
transaction type: SELECT only
scaling factor: 100
query mode: simple
number of clients: 64
number of threads: 64
duration: 300 s
number of transactions actually processed: 71851589
tps = 239503.585813 (including connections establishing)
tps = 239521.816698 (excluding connections establishing)
real 5m0.016s
user 20m40.880s
sys 9m25.930s
Since a client-limited benchmark isn't very interesting, I think this
change makes sense. Thoughts? Objections? Coding style
improvements?
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company