Ants Aasma <ants.aasma@eesti.ee> writes:
> On Mon, Aug 29, 2011 at 10:12 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Also, if the PPC machine really is hyperthreaded (the internal webpage
>> for it says "Hyper? True" but /proc/cpuinfo doesn't provide any clear
>> indications), that might mean it's not going to scale too well past 16x
>> the single-thread case.
> According to IBM docs [1], 8406-71Y contains one 8 core POWER7 chip
> that is 4-way multi-threaded and has 4 memory channels.
Yeah, I looked at the docs. "Multi threading" is IBM's term for
hyperthreading, that is several instruction streams competing for use of
a single processor core's pipelines. So the 32 virtual processors on
the machine only really represent 8 physically independent cores,
squaring with the hardware designation. I found an IBM doc
http://www-03.ibm.com/systems/resources/pwrsysperf_SMT4OnP7.pdf
suggesting that the throughput benefit of 4-way SMT is typically 1.5 to
2X, that is you max out at 1.5 to 2X as much work as you'd get with just
8 virtual processors on the same 8 cores. So I'd say we're really doing
quite well to get the numbers I got. (The paper also implies that you
get more benefit from SMT with workloads that incur more memory-access
stalls, so the relatively large working set of this test case is helping
it look good.)
> 6GB of
> memory is a strange amount for the IBM, according to specs it should take
> 4 or 8GB DIMMs in pairs. Sounds like the server is split into multiple
> partitions.
I'm confused about that too. There definitely seemed to be only 6GB of
available RAM, but there's no way I can see that memory might be
partitioned across different blades. The blades look pretty independent
...
regards, tom lane
