[ale] Hyperthreading

[Jeff Layton]

It generally hurts our CFD codes. It also hurts all but one of the
NPB codes (which are CFD). We turn it off for now.

Maybe in the future when a real multi-core chip comes out
we can revisit it. Multi-core chips also give rise to the possibility
of doing multi-level parallelism usch as MPI+OpenMP.

Jeff

>In /proc/cpuinfo, you'll see an "HT" flag.  Supposedly that indicates
>HyperThreading support on the CPU.  In the case of my workstation, the
>BIOS doesn't support HT, but the CPU reports that it does, so I'm out of
>luck.
>
>In our CFD (computational fluid dynamics) research, we've seen a wide
>range of speed changes using HT.  Some are great, and others are
>deplorable.  I think it mainly has to do with the APP and the data being
>crunched.
>
>Thanks
>
>Jonathan Glass
>
>On Wed, 2004-04-14 at 14:00, Steven A. DuChene wrote:
>  
>
>>I have been reading the responses in this thread so far but i have not as of
>>yet actually heard anything about checking to see if a particular P4 processor
>>is really Hyperthread capable. Not all P4 processors are created equal and
>>some are not Hyperthread capable. Typically other than actually research the
>>particular processor specs on the Intel web site the best way to tell is to
>>see if there is an option in the motherboard BIOS to turn Hyperthreading on
>>or off. Once it is turned on if you then run a SMP kernel you will actually
>>see more than one processor being reported in /proc/cpuinfo
>>
>>If you run an SMP kernel and still only see one processor in /proc/cpuinfo
>>then either the Hyperthreading support is not turned on OR your processor
>>is not Hyperthread capable.
>>
>>Now, as to whether it is a good thing or a bad thing to turn on, it is all
>>a question of the particular types of work loads the system sees. There are
>>cases where Hyperthreading is not actually a help and in some cases it may
>>actually be a performance degradation.
>>
>>On Wed, Apr 14, 2004 at 10:40:40AM -0400, Dow Hurst wrote:
>>    
>>
>>>Yes, I run two minimizations simultaneously on my single P4 and until the 
>>>second one starts the machine is very responsive.  It is truly like having a 
>>>dual CPU from my user perspective.  Run times are doubled on the minimizations.
>>>
>>>25m37.821s   Single run
>>>
>>>49m54.143s
>>>44m10.795s   Dual runs (The second is the shorter due to different start 
>>>times, so this is general datapoints)
>>>
>>>
>>>These are GPCR receptor minimizations with delta 9 THC bound inside the 
>>>receptor.  About ~2700 atoms in vacuum using molecular mechanics and a lot of 
>>>torsional restraints applied.  Not a difficult calculation but is pure FPU on 
>>>C and Fortran code.  But, my point is the machine was acting like a dual CPU 
>>>machine in a very efficient manner.
>>>Dow
>>>
>>>
>>>Christopher Fowler wrote:
>>>      
>>>
>>>>How is the performance? Is there a real gain?
>>>>
>>>>On Wed, 2004-04-14 at 10:03, Jonathan Glass wrote:
>>>>
>>>>        
>>>>
>>>>>On Wed, 2004-04-14 at 10:02, Mike Millson wrote:
>>>>>
>>>>>          
>>>>>
>>>>>>I have a box with a P4 processor running Slackware 9.1. What do I need
>>>>>>to do to enable hyperthreading support? Do I just have to compile a
>>>>>>kernel with SMP support? Are there any other kernel options or config
>>>>>>changes I need to make? Does anyone know of any good articles or howtos
>>>>>>on configuring linux for hyperthreading?
>>>>>>
>>>>>>Thank you,
>>>>>>Mike
>>>>>>
>>>>>>            
>>>>>>


-- 
Dr. Jeff Layton
Aerodynamics and CFD
Lockheed-Martin Aeronautical Company - Marietta