PC GAMESS/Firefly-related discussion club

Learn how to ask questions correctly

Alex Granovsky
gran@classic.chem.msu.su

we hope to publish data on Core i7 with HTT enabled within a week or so, as we already have results of our benchmarks.
Next, within 2-3 weeks the data on Xeons 5500 will be published here as well. Once again, we have these data for at least a half of year,
however they were obtained on engineering samples and thus cannot be published. We need to repeat tests on production level systems first.


In brief, there is typically a speedup of ten to twenty percents when
using all available logical cores with Firefly, as compared with runs
using single logical core per physical core. However, speedup varies
depending on the particular job. Other programs that are not HTT
aware may have degraded performance running with HTT enabled.

I'll post more information to this thread when the results of benchmarks will be published in the "Performance" section.

Regards,
Alex



On Mon Apr 13 '09 8:03pm, Veinardi Suendo wrote
-----------------------------------------------
>Dear Alex and Firefly Users,

>Actually, we are planning to expand our system to this new processor family due to its HTT. However, does this function really reduce the calculation performance? I have checked in the performance section but I could not find the benchmark that used all 8 cores (4 real cores + 4 logical cores). Is it comparable the performance of one Core i7 machine with 2 Core2 machines? If you do not mind would you explain more about Firefly options using HTT. Normally, in Core 2 Quad based machine, I use only:

> $P2P p2p=.t. dlb=.t. xdlb=.t. mxbuf=2048 $END
> $SMP Call64=.t. httpar=.t. httalt=.t. $END

>How about in Core i7 machine? Should I add more option in $SMP section such the following?

>$SMP Call64=.t. httpar=.t. httalt=.t. httfix=.t. $END

>Thank you very much in advance,

>Best regards,

>Veinardi

>On Wed Apr 8 '09 0:27am, Alex Granovsky wrote
>---------------------------------------------
>>Dear PC GAMESS/Firefly users,

>>Recent Intel's Nehalem core processors (e.g., Core i7/Xeon 5500/Xeon 3500)
>>reimplement hyper-threading technology, which does result in some performance improvements
>>(see results of our tests in the "Performance" section).

>>Unfortunately, as of moment of writing this note (April 2009),
>>both Windows and Linux are not capable to identify correctly
>>the particular logical processors belonging to each specific CPU
>>core. In practice this confuses tasks/threads scheduler and
>>results in performance degradation of parallel or multithreaded jobs
>>while running with HTT turned on.

>>This was found for Windows Vista/Windows Server 2008 and most likely
>>affects older versions of Windows as well. As of now, all modern
>>Linux distributions with their default kernels are also affected.
>>
>>
>>The reason of this is most likely the changed semantics of some CPUID functions.

>>This problem, however, does not affect Firefly running with httfix
>>option turned on (see http://classic.chem.msu.su/gran/gamess/smp.html for details on this option).
>>
>>
>>This option is turned on under Windows by default; however, to use it
>>under Linux, one needs to upgrade to Firefly version 7.1.E or above
>>and turn it on manually.
>>
>>
>>In addition, one may need to provide correct -ncores and -nthreads command line arguments
>>(see http://classic.chem.msu.su/gran/gamess/comm_line.html for details),
>>e.g. if running Firefly in environment with CPUID instruction virtualized.
>>For example, MS Hyper-V (and perhaps some other modern virtualization solutions)
>>virtualizes CPUID instruction thus limiting the amount of information available to Firefly.
>>This affects Firefly running both in the virtual machines and on the host OS itself.

>>Alternatively, one may disable HTT at all. As to Firefly running
>>using correct options described here, this will result in some
>>performance degradation. However, this may improve performance of
>>other programs one may need to use.

>>Best regards,
>>Alex Granovsky
>>
>>
>>

[ This message was edited on Sat May 2 '09 at 12:04pm by the author ]