comp.lang.idl-pvwave archive: archive » Re: Help, no improvement in FFT speed on a multiprocessor system

Home » Public Forums » archive » Re: Help, no improvement in FFT speed on a multiprocessor system

Show: Today's Messages :: Show Polls :: Message Navigator
E-mail to friend

Re: Help, no improvement in FFT speed on a multiprocessor system [message #67897]

Mon, 07 September 2009 14:30

Kenneth P. Bowman
Messages: 585
Registered: May 2000

Senior Member

In article <4aa56605$1@darkstar>, "Marco" <null@null.net> wrote:

> The arrays are Nx8192 on a side with N a power of 2.
>
> I increased N until the speed dropped of a cliff. Presumably cache/memory
> thrashing.
>
>
> "Kenneth P. Bowman" <k-bowman@null.edu> wrote in message
> news:k-bowman-5D1FE7.08403607092009@news.tamu.edu...
>> In article <4aa34391$1@darkstar>, "Marco" <null@null.net> wrote:
>>
>>> I'm running IDL 7.1 on a Linux 2.6. This is an HP quad processor with
>>> each
>>> processors having 6 cores for 24 cores total.
>>>
>>> Doing large 2-D FFTs (>8Kx8K) I get no benefit from the extra
>>> processors.

I don't know how the Intel cache architecture works, but on some
processors (e.g., IBM Power), a cache miss causes a whole cache
line to be loaded from memory. If you are working on large arrays
and taking large strides through memory, every memory access
can cause a cache miss. This has the effect of completely
destroying the advantages of having a cache. Arrays dimensioned
by powers of two can be the worst cases.

I don't know an easy solution. You could do N 1-D FFTs of size
8192, transpose the output, and then do 8192 1-D FFTs of size N.
That is, "manually" make a 2-D FFT by looping over the second
dimension. It might possibly be faster than doing a 2-D FFT with
miserable cache performance.

Ken Bowman

Report message to a moderator

[Message index]

		Re: Help, no improvement in FFT speed on a multiprocessor system By: Kenneth P. Bowman on Mon, 07 September 2009 14:30
		Re: Help, no improvement in FFT speed on a multiprocessor system By: Marco on Mon, 07 September 2009 12:58
		Re: Help, no improvement in FFT speed on a multiprocessor system By: Kenneth P. Bowman on Mon, 07 September 2009 06:40

Previous Topic:	Re: Function to return the current procedure/function
Next Topic:	Re: use of loops in IDL for hashtable

-=] Back to Top [=-

[ Syndicate this forum (XML) ] [

] [

]

Current Time: Sun Nov 30 01:37:32 PST 2025

Total time taken to generate the page: 1.03985 seconds