[phenixbb] Optimal Linux configuration for Phenix.refine
tg at shelx.uni-ac.gwdg.de
Wed Jun 24 11:47:35 PDT 2015
I would think that using the virtual cores from hyperthreading slows
down the overall wall-clock time and would use the number of real cores,
maybe +1 in case there are steps in the program that don't fully occupy
Otherwise I also have the impression that i7 CPUs are significantly
faster than Xeon machines with comparable clock frequency.
On 06/24/2015 07:32 PM, Jim Fairman wrote:
> I can't comment much on refinement with phenix.den as it is a feature I
> haven't used before.
> There is indeed an input for number of processors. If you are running via
> command line, add --nproc=# to your command, where # is the number of
> processors. If you are using the GUI, there is a box in the bottom
> right-hand corner where you can select number of CPUs.
> For hyper threaded CPUs you can pick double the actual physical cores. The
> 4770 chips have 4 hyper threaded physical cores, so you can run jobs with
> up to eight.
> On Wed, Jun 24, 2015 at 10:21 AM, Xiao Lei <xiaoleiusc at gmail.com> wrote:
>> Hi Jim,
>> Thanks for the information! I just got a HP with core i7 4771 (4 cores
>> 3.5GHz), now I know it will be a good speed boost for phenix. I used DEN
>> refinement in phenix, which is very slow, I'm even thinking using
>> university cluster to do it. There is an "DEN parameter optimization"
>> choice in phenix, I do not know if this can be finished by parallel
>> computing with multiple processors.
>> I have one question, in the phenix refinement, there is an input for the
>> number of processor, is it the same number as cores CPU has? If a CPU is
>> hyper-threaded like core i7 4771 (8 threads but 4 cores), should I put
>> number of processors 8 or 4?
>> On Wed, Jun 24, 2015 at 9:06 AM, Jim Fairman <fairman.jim at gmail.com>
>>> A few years back one of the PHENIX devs referenced Amdahl's Law and how
>>> it affects the performance of PHENIX:
>>> Even though some parts of the refinement are done in parallel on multiple
>>> processors, the really time-consuming calculations are done in serial,
>>> which ends up governing the overall speed of the process.
>>> I have found that Intel Haswell-class processors like the Core i7 4770
>>> perform refinements significantly speedier than our brand new 12-core Xeon
>>> server that is less than 6 months old. So it is my naive guess that
>>> getting the fastest "single processor performance" will get you maximal
>>> refinement speeds on PHENIX.
>>> If you want to up the ante a bit more, you can always try overclocking
>>> with liquid nitrogen to hit 7.2 Ghz <http://valid.canardpc.com/240rw0>.
>>> Cheers, Jim
>>> On Wed, Jun 24, 2015 at 8:20 AM, Colin Levy <C.Levy at manchester.ac.uk>
>>>> Apologies for the slightly off topic question.
>>>> I am configuring a Linux workstation and would like your thoughts on
>>>> an optimal setup for running Phenix. A lot of my computing time is spent in
>>>> refinement and I am keen to put something together that will see a
>>>> considerable increase in speed from my creaking Mac Pro (2 2.4GHz Quad core
>>>> Intel Xeon with 24Gb RAM).
>>>> Many thanks,
>>>> Dr. Colin W. Levy
>>>> MIB G034
>>>> Tel. 0161 275 5090
>>>> Mob.07786 197 554
>>>> c.levy at manchester.ac.uk
>>>> phenixbb mailing list
>>>> phenixbb at phenix-online.org
>>>> Unsubscribe: phenixbb-leave at phenix-online.org
>>> Jim Fairman, Ph D.
>>> Group Leader I - Crystallography
>>> Beryllium <http://www.be4.com>
>>> Tel: 206-780-8914
>>> Cell: 240-479-6575
>>> E-mail: fairman.jim at gmail.com jfairman at be4.com
>>> phenixbb mailing list
>>> phenixbb at phenix-online.org
>>> Unsubscribe: phenixbb-leave at phenix-online.org
> phenixbb mailing list
> phenixbb at phenix-online.org
> Unsubscribe: phenixbb-leave at phenix-online.org
Dr Tim Gruene
Institut fuer anorganische Chemie
phone: +49 (0)551 39 22149
GPG Key ID = A46BEE1A
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 173 bytes
Desc: OpenPGP digital signature
More information about the phenixbb