On Mon, Aug 29, 2011 at 2:15 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> These tests were run on a 32-CPU Opteron machine (Sun Fire X4600 M2,
> 8 quad-core sockets). Test conditions the same as my IA64 set, except
> for the OS and the -j switches:
>
> Stock git head:
>
> pgbench -c 1 -j 1 -S -T 300 bench tps = 9515.435401 (including ...
> pgbench -c 32 -j 16 -S -T 300 bench tps = 227349.424654 (including ...
>
> These tests were run on a 32-processor PPC64 machine (IBM 8406-71Y,
> POWER7 architecture; I think it might be 16 cores with hyperthreading,
> not sure). The machine has "only" 6GB of RAM so I set shared_buffers to
> 4GB, other test conditions the same:
>
> Stock git head:
>
> pgbench -c 1 -j 1 -S -T 300 bench tps = 8746.076443 (including ...
> pgbench -c 32 -j 16 -S -T 300 bench tps = 126822.857978 (including ...
Stepping beyond the immediate issue of whether we want an unlocked
test in there or not (and I agree that based on these numbers we
don't), there's a clear and puzzling difference between those sets of
numbers. The Opteron test is showing 32 clients getting about 23.9
times the throughput of a single client, which is not exactly linear
but is at least respectable, whereas the PPC64 test is showing 32
clients getting just 14.5 times the throughput of a single client,
which is pretty significantly less good. Moreover, cranking it up to
64 clients is squeezing a significant amount of additional work out on
Opteron, but not on PPC64. The HP-UX/Itanium numbers in my OP give a
ratio of 17.3x - a little better than your PPC64 numbers, but
certainly not great.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
