On 10/03/08 13:20, Elad Lahav wrote:
>> (In your first posting you said mpstat was used, not vmstat, so
>> I will assume mpstat).
> Oops, of course, I meant mpstat...
>
>> If mpstat shows idle time, then during the idle time, no thread
>> is runnable from a high-level, traditional operating system point
>> of view.
> Not sure about that. mpstat shows idle time for a virtual processor,
> representing a single hardware strand. Are you suggesting that this
> implies that all 4 threads on the core are idle?
I meant: if mpstat shows idle time for a CPU, then no thread is
runnable for that CPU during the idle time. In your example, no
other threads are running on the core, so mpstat must look like this:
% mpstat
CPU usr+sys idle
0 50% 50% \
1 0% 100% \ these 4 CPUS share a core pipeline
2 0% 100% /
3 0% 100% /
...
hence 50% of the time, there are no runnable software threads for
the core and all 4 strands are idle.
In case we are using different terminology, here is mine:
Solaris CPU == virtual processor == hardware strand
mpstat shows each CPU individually.
There are 4 CPUs per core, which share an execution pipeline (on the T1000).
The OS assigns a software thread to a CPU when the software thread
becomes runnable.
- Steve
>
>> In your example, mpstat showed a 50% busy server thread, and
>> a 50% busy interrupt CPU, on different cores. Maybe the critical
>> path runs from the interrupt thread to to app thread with no
>> overlap, hence each can only stay half busy.
> Perhaps.
>
> > Or maybe you have a bottleneck on the client side.
> No. Substituting the T1000 for a Xeon server yields a 1 Gbps throughput,
> so it's definitely not the client's fault.
>
>> Also let me add the standard CMT advice: CMT is designed for
>> aggregate thruput and gives the best performance when many threads
>> run concurrently. Your single threaded test is not the best example
>> for this architecture.
> I am well aware of that. I am only trying to figure out the capabilities
> of strands and cores, and the effect of different placement policies, in
> order to come up with good ways of parallelising network servers.
>
> At this point I am inclined to believe it has something to do with TCP,
> as alluded to by David in a previous response. I have a similar problem
> when using Linux, and there moving to UDP yields much better results. I
> will give it a try on Solaris now.
>
> Thanks for your help,
> --Elad
>
_______________________________________________
perf-discuss mailing list
perf-discuss@opensolaris.org
