Alex Granovsky
gran@classic.chem.msu.su
>this question is really important, at least for every Firefly
>users running it under Windows 7 on Nehalem processors (i.e.,
>Core i7, Xeon 3500, Xeon 5500 etc...).
>Actually, all the retail versions of Windows 7 we tested so far
>(we did not test Win2K8 R2 as of yet) have the nasty kernel bug
>that causes really weird things to happen if the process has an
>explicit CPU affinity mask set. This causes dramatic performance
>degradation if one is running PC GAMESS or Firefly in parallel on
>e.g. Core i7 system with hyper-threading enabled. There are two
>workarounds available at moment. The first one is to disable
>hyperthreading in BIOS. The second one is to set $smp httfix=0 smppar=1 $end
>and always run Firefly in parallel using all available logical,
>not physical cores. The latter would mean that you may need up
>to 8 physical disks (or a couple of SSDs) for disk-intensive tasks
>like MP2 code.
>Another workaround is to resort to Windows Vista or Win2K8 R1...
>We are working with MS to resolve this issue. Actually,
>Windows kernel team has been notified on this problem and
>most likely MS will provide a fix to this problem soon enough.
>Hope this helps.
>Regards,
>Alex Granovsky
>
>
>On Tue Dec 29 '09 1:21am, Fa-Gung Fan wrote
>-------------------------------------------
>>Dear Firefly/PC GAMESS Users:
>>I recently obtained a Core i7 920 system, and opted to use Windows 7 Enterprise edition 64-bit. As soon as the system was up, I read this post of Dr. Monev regarding the performance of Windows 7 vs. Vista for Firefly/PC GAMESS calculations. I thought “Oops, I might have made a mistake, should have selected Vista instead.”
>>I would like to ask on this forum, since Dr. Monev’s experience with Windows 7 RC in August, has someone investigated the OS’ performance (especially, the released editions) for Firefly/PC GAMESS? Please post your experiences and suggestions. Thank you very much.
>>
>>
>> Fa-Gung Fan
>>
>>
>>On Sat Aug 1 '09 3:12pm, Valentin Monev wrote
>>---------------------------------------------
>>>Dear Professor Granovsky,
>>>I compared wall times for mp2 optimization runs for two identical pc’s (core i7 cpu 920 6GB ram 2x1TB hdds, hyperthreading and turbo-boost enabled) under windows 7 (RC) and vista (both ultimate, 64 bit). To my surprise, the windows 7 wall times were twice or more greater than the Vista wall times – see attached files (1 input, 2 output files). In addition, running the same input in a row under windows 7, I obtained wildly different wall times. To exclude any hardware problems, I replaced 7 with vista, and the wall times became stable and comparable to those for vista on the other pc. I reinstalled windows 7, this time with the 2 hard disks in raid0 (intel ich10r, enabled write caching, turned off windows write-cache buffer flushing), for additional convenience and performance benefits. Again, wall times became unstable and much longer than those for the vista system.
>>>As advised by you (http://classic.chem.msu.su/cgi-bin/ceilidh.exe/gran/gamess/forum/?C34df668afbHW-7096-1266+00.htm), I used a single fixed logical processor per core, and the default $smp values (see attached files). [Previously, I had tried using 7 logical processors out of the total 8 available ($smp httfix=.f. httpar=.t. smppar=.t. $end.) In fact, this was often much faster per point, but overall the gradients tended to decrease faster i.e. final results were obtained faster in the case of just 1 logical processor per core (4 fixed logical processors). Additional benefits of 4 fixed logical processors: a) the pc is loaded just 50nstead of 89řabout 8 degrees centigrade cooler, and the pc is more responsive for interactive work), b) in (next) version G of firefly, “SMT enabled parts of the program” (http://classic.chem.msu.su/cgi-bin/ceilidh.exe/gran/gamess/forum/?C34df668afbHW-7096-1268+00.htm) will further utilize the second (free) logical processor.]
>>>I tried many combinations of options described in the (new) mp2 documentation. For example, leaving out the line $mp2 ioflgs(1)=65536,65536 ioflgs(3)=65536,65536 $end led to the memory becoming occupied 99°and the pc became totally unfit for interactive work, while the performance issue was even aggravated. Leaving out XDLB=.T., and/or mxbuf=2048, did not help either. Restarting the pc sometimes seemed to worsen the performance, which is strange.
>>>To summarize, I have no idea what is different in windows 7 relative to vista, but it seems tuning of firefly may be needed for the new operating system.
>>>Of course I realize that it may be asking too much for you to tune the program for an operating system that is yet to come out in October 2009. But I could run the program with various options suggested by you…
>>>Thank you in advance.
>>>