Print

Print


2009/2/18 Mark Taylor <[log in to unmask]>:
> On Wed, 18 Feb 2009, David Berry wrote:
>
>> 2009/2/18 David Berry <[log in to unmask]>:
>> > 2009/2/18 Mark Taylor <[log in to unmask]>:
>> >> Peter, David et al.,
>> >>
>> >> The JNIAST version on the starjava svn trunk now works with AST threads,
>> >> including allowing multiple AST calls to execute concurrently as long
>> >> as the same objects are not involved, and as long as AST has been
>> >> compiled --with-pthreads.  I believe there are no outstanding threading
>> >> issues, but let me know if anyone spots something that looks like one.
>> >> My rudimentary benchmarking suggests that scaling is not very good
>> >> (elapsed time for a 2-core processor, using 100% of both, is nearly
>> >> the same as for the same thing running without threads enabled on
>> >> a single core), but that's a different issue.
>> >
>> > I'll do some timing tests on smurf to check the effect of the recent
>> > fixes I have made to AST in the threading area.
>>
>>
>> On 28th October, in mail thread "Local directory on sc2dr1", I said:
>>
>>
>> > The overall cpu usage is 475%, including the non-threaded parts.
>> >
>> > This was using the command
>> >
>> > % time $SMURF_DIR/makecube a20070729_00015_01_0001 spread=gauss
>> > params=\[2,2\] out=fred-test msg_filter=verb
>>
>> Using the same version of smurf but the current version of AST, this
>> command now runs at an overall cpu usage of 458 %. So a bit slower for
>> some reason, but nothing to write home about.
>
> To clarify: I haven't had any problems getting high CPU usage in my
> (short) experiments, only with seeing the benefits from this,
> i.e. seeing reduced elapsed times for multi-core operation.  However
> it's quite possible that my results are to do with the interactions
> between java and C threading, or would go away if I did more
> comprehensive benchmarking, or something else.  I'm not currently
> planning to spend more effort investigating this.

I did the tests just to check that the changes I've made recently to
thread support in AST have not had an adverse effect on smurf. The
fact that I'm seeing an overall cpu of 450%, peaking at around 650%,
on an 8 core system, suggests that the threading still seems to be
working reasonably in smurf. That is, running the same command on a
single core (i.e. cpu<=100%) would take at least 4.5 times longer than
on the 8 core system.

David