With the processor allocation bolted down, the run across eight Opterons with three threads each took real-time 50757 seconds.

With 24 MPI tasks at one thread each it's reporting ETA 16 hours 6 minutes at 16.1%, so about 70k seconds ... so the Haswell is going to be faster. This is really not at all what I expected. The MPI x24 has speeded up a little when I stop every other task on the machine,

I was seeing CPU usage in the low-200% range for each of the eight processes in the multi-threaded case; with 24 MPI tasks, each is at precisely 100%.

Maybe try a 2x4 or 4x2 grid? The aspect ratio has a big effect on the total amount of communication, as well as how much of it can be overlapped.

Also check that the choices of block sizes make sense; you want the block size to be about the L1 size and the superblock size to be maybe 3/4 the size of the largest cache a CPU can use, and can override the default with 'la_block=X' and 'la_superblock=Y' (with X and Y the number of 8-byte words). Also, if the AMD processors don't have automatic prefetch, we can add manual prefetches to the code in one or two places, to see if that makes a difference. If the utilization is only 200% with three threads, then two of those threads will be idle fairly often wating for the thrid to finish it's parts of a multithreaded matrix multiply; it would possibly be better to use just two threads.

24-thread run would have taken 52782 seconds if running unhindered.

I've set 2x4 and 4x2 grid -t2 and -t3 runs to go while I'm on vacation next week.

Should the superblock size be the size of the L2 cache shared by cores 0-2, or one third of the size of that cache?

The new -nc2 code seems blazing fast in rev. 900. Thanks!

It seems like -nc1 is a bit slower, though; is that expected?

Filtering should not have any changes...

The superblock size should be almost the whole size of the cache that will be shared across all the threads.

Could anyone on Windows 64 bit compile SVN 900 for me?
 

Would it be possible for Brian Goldman or Jeff Gilchrist or anyone that feels confident to compile me a copy of 1.52 pre-release with SVN 900?

My system is a I 7 980 X running at 3.43 GHz with with 12 gig of RAM. Operating system Windows 7 ultimate fully patched 64 bit If you need any more details please ask.

Very grateful for any help

Thanks
Jarod

[QUOTE=Speedy51;345039]Would it be possible for Brian Goldman or Jeff Gilchrist or anyone that feels confident to compile me a copy of 1.52 pre-release with SVN 900?

My system is a I 7 980 X running at 3.43 GHz with with 12 gig of RAM. Operating system Windows 7 ultimate fully patched 64 bit If you need any more details please ask.

Very grateful for any help

Thanks
Jarod[/QUOTE]

I am guessing that you mean me.

I have only just found out that my builld is out of date because the SourceForge update has left me on the old repository.

But I have now built the latest (non GPU) version, which I can make available if Jeff is willing to host it.

I am still unable to build the GPU version because NVIDIA has been amazingly lazy in not updating their tools for Visual Studio 2012. Amazingly we still don't have any NVIDIA 2012 support even though we now have a pre-release of Visual Studio 2013!

It appears that Microsoft and Nvidia are now pushing C++ AMP as the way to harness GPU computing on the Windows platform, which may impact on building the GPU code with the Microsoft tools.

[QUOTE=Brian Gladman;345059]I am guessing that you mean me.

I have only just found out that my builld is out of date because the SourceForge update has left me on the old repository.

But I have now built the latest (non GPU) version, which I can make available if Jeff is willing to host it.

I am still unable to build the GPU version because NVIDIA has been amazingly lazy in not updating their tools for Visual Studio 2012. Amazingly we still don't have any NVIDIA 2012 support even though we now have a pre-release of Visual Studio 2013!

It appears that Microsoft and Nvidia are now pushing C++ AMP as the way to harness GPU computing on the Windows platform, which may impact on building the GPU code with the Microsoft tools.[/QUOTE]
Yes Brian you are correct, sorry for getting your last name wrong. The future of GPU at certainly does look interesting I hope they can get it sorted out. I look forward to waiting to hear whether Jeff will host your latest build. I'm looking forward to seeing how much quicker there SVN 900 code is.

Thanks to Brian for sending me a compiled version, you can now download the Windows 64bit version of SVN902 here:
[url]http://gilchrist.ca/jeff/factoring/msieve152_svn902_win64_i7.zip[/url]

A DLL file maybe missing?
 

[QUOTE=Jeff Gilchrist;345116]Thanks to Brian for sending me a compiled version, you can now download the Windows 64bit version of SVN902 here:
[URL]http://gilchrist.ca/jeff/factoring/msieve152_svn902_win64_i7.zip[/URL][/QUOTE]
Thanks for posting Jeff. I noticed there is no large block application? I unzipped the files into my working directory and changed the command line to read as follows [code]start /low /min msieve.exe -i L1213.ini -s L1213.dat -nf L1213.fb -v -ncr -nc2 -t 3 [/code] I used to use [code]start /low /min msieve.largeblocks.exe -i L1213.ini -s L1213.dat -nf L1213.fb -v -ncr -nc3 -t 3[/code]when I double clicked the bat file I get The following error [code]the program can't start because VCOMP110.DLL is missing from your computer. Try reinstalling the program to fix this problem. [/code] if I use the large block application the LA process restarts fine but it uses version 1.51 official release. I double checked in the zipped file and I cannot see the DLL file mentioned above

A separate largeblocks compile is no longer needed. Code changes use the appropriate settings for your CPU.

I think the vcomp110.dll is part of the Visual C++ 2012 redistributable here:
[url]http://www.microsoft.com/en-us/download/details.aspx?id=30679[/url]