23.7 - not much of interest

Prime95 · 2003-09-06, 20:27

V23.7 contains a torture test bug fix. Some users were reporting "Out of memory" errors and excessive virtual memory usage.

ewmayer · 2003-09-07, 00:50

Just out of curiosity, what was the bug and how did you fix it? I've run into VM issues occasionally when doing big sets of self-tests with Mlucas - my planned fix is to allocate (say) twice as much memory as a given LL test needs, then re-use the same memory space until the self-test exponent gets so large that it needs more than that, at which point I'll realloc to twice the space. That will reduce the program's dependence on the OS's VM management capabilities during self-tests.

Prime95 · 2003-09-07, 01:39

That's exactly what I did. In 23.6, if doing a 300MB torture test, I'd malloc the FFT areas one at a time, free them when done, malloc again for the next exponent. In 23.7 I allocate one 300MB buffer up front that isn't freed until the torture test ends and wrote a routine to carve each FFT area out of this big buffer.

Xyzzy · 2003-09-07, 04:12

I thought something was up...

I set Prime95 to use 1841MB for a torture test... Out of 2048 that was the max... Talk about some disk thrashing!

:(

ewmayer · 2003-09-08, 15:29

Yes, re-using the same single memory space for the entire self-test suite ahs the advantage that it would reduce or eliminate the influence of cache coloring effects on timings. Since Mlucas' maximum FFT length (8M) needs just 64MB of memory, preallocating that much for the entire self-test (or when running regular LL tests) wouldn't be a problem.

xtreme2k · 2003-09-24, 15:34

The "Blend" option of the torture test takes too much memory by default.

I have 2GBs of RAM and it takes 1951MB by default causing windows to swap everything into the swap file back and forth instead of actually stressing the CPU.

Also what exactly does 'run FFTs in-place' mean?

Xyzzy · 2003-09-24, 15:38

Under Linux I can allocate 2017MB out of 2048!

(Not a good idea, BTW!)

PS - Using 23.5, there are no newer Linux versions...

2003-09-06, 20:27	#1
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 2×4,079 Posts	23.7 - not much of interest V23.7 contains a torture test bug fix. Some users were reporting "Out of memory" errors and excessive virtual memory usage.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Would there be any interest in another subproject?	fivemack	Aliquot Sequences	20	2013-01-29 11:10
Bootable USB/CD for Sequences - Any Interest?	EdH	Aliquot Sequences	0	2010-02-20 15:33
Interest/participation seems to be up again!	petrw1	PrimeNet	1	2007-11-27 17:27
this of interest to anyone?	ixfd64	Hardware	5	2005-11-14 01:44
Any interest in some low n ECM factoring?	ltd	Prime Sierpinski Project	19	2005-08-12 06:24

2003-09-07, 00:50	#2
ewmayer ∂²ω=0 Sep 2002 República de California 10110111101011₂ Posts	Just out of curiosity, what was the bug and how did you fix it? I've run into VM issues occasionally when doing big sets of self-tests with Mlucas - my planned fix is to allocate (say) twice as much memory as a given LL test needs, then re-use the same memory space until the self-test exponent gets so large that it needs more than that, at which point I'll realloc to twice the space. That will reduce the program's dependence on the OS's VM management capabilities during self-tests.

2003-09-07, 01:39	#3
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 1FDE₁₆ Posts	That's exactly what I did. In 23.6, if doing a 300MB torture test, I'd malloc the FFT areas one at a time, free them when done, malloc again for the next exponent. In 23.7 I allocate one 300MB buffer up front that isn't freed until the torture test ends and wrote a routine to carve each FFT area out of this big buffer.

2003-09-07, 04:12	#4
Xyzzy Aug 2002 2×7×13×47 Posts	I thought something was up... I set Prime95 to use 1841MB for a torture test... Out of 2048 that was the max... Talk about some disk thrashing! :(

2003-09-08, 15:29	#5
ewmayer ∂²ω=0 Sep 2002 República de California 5×2,351 Posts	Yes, re-using the same single memory space for the entire self-test suite ahs the advantage that it would reduce or eliminate the influence of cache coloring effects on timings. Since Mlucas' maximum FFT length (8M) needs just 64MB of memory, preallocating that much for the entire self-test (or when running regular LL tests) wouldn't be a problem.

2003-09-24, 15:34	#6
xtreme2k Aug 2002 2×3×29 Posts	The "Blend" option of the torture test takes too much memory by default. I have 2GBs of RAM and it takes 1951MB by default causing windows to swap everything into the swap file back and forth instead of actually stressing the CPU. Also what exactly does 'run FFTs in-place' mean?

2003-09-24, 15:38	#7
Xyzzy Aug 2002 216A₁₆ Posts	Under Linux I can allocate 2017MB out of 2048! (Not a good idea, BTW!) PS - Using 23.5, there are no newer Linux versions...