| Re: When does IDL use MPs? [message #32888 is a reply to message #32843] |
Fri, 15 November 2002 14:20  |
JD Smith
Messages: 850
Registered: December 1999
|
Senior Member |
|
|
On Wed, 13 Nov 2002 13:02:45 -0700, David Fanning wrote:
> Jonathan Greenberg (greenberg@ucdavis.edu) writes:
>
>> Do I need to specifically invoke multiprocessor calls in an IDL
>> procedure, or will it use MPs whenever appropriate?
>
> Oh, oh. Those marketing guys have gotten out again. :-(
>
> I'm afraid IDL is not multi-threaded. It can use multiprocessors in one
> very narrowly defined situation: when it is rendering object graphics
> volumes. Other than that, you can invoke until the cows come home, but
> it's not going to help much. Prayer would be more effective, I think.
> :-)
>
Well, it's slightly better than that. IDL is designed to use its
thread pool for a number of the most important operations and routines
(WHERE, FFT, basic array arithmetic, etc.) on MP machines whenever
appropriate. The bad news is that the overhead for spawning a
separate thread to compute, say, the total of a big array is
significant, motivating the default choice for the minimum number of
elements before threading occurs to be large:
help,!CPU,/STRUCTURES
** Structure !CPU, 6 tags, length=24, data length=24:
HW_VECTOR LONG 0
VECTOR_ENABLE LONG 0
HW_NCPU LONG 2
TPOOL_NTHREADS LONG 2
TPOOL_MIN_ELTS LONG 100000
TPOOL_MAX_ELTS LONG 0
I.e. unless you're operating on an array with more than 100,000
elements, you are out of luck (TPOOOL_MIN_ELTS). You can change this
value globally with the CPU procedure, or hand tune it in the call to
any of the 50 or so routines which are threaded, but 100,000 is
usually about right.
When the magic is right, it really cooks on my dual-processor system:
IDL> r=randomu(sd,10000000L)
IDL> t=systime(1) & tot=total(r,TPOOL_MIN_ELTS=10000001L) & print,systime(1)-t
0.11507905
IDL> t=systime(1) & tot=total(r,TPOOL_MIN_ELTS=10000000L) & print,systime(1)-t
0.061390042
I find measurable speedups even for 50,000 elements, but it really
shines on much larger chunks of data. It also depends on the
function, of course. Experiment and see what you find.
JD
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
