[Beowulf] Threaded code

Art Edwards arthur.edwards at kirtland.af.mil
Tue Aug 17 13:30:56 EDT 2004


Only a tuned blas3 increases the speed, so I'm talking about a serial
version of atlas tuned on a specific node. The code we have been working
with is actually a highly tuned serial code. On a single node, it
performs 20 times faster than other parallel codes. We are only now
looking at threading and at MPI parallelization.

Art

On Tue, Aug 17, 2004 at 01:26:21PM -0400, Joe Landman wrote:
> Art Edwards wrote:
> 
> >It is conceivable, but unlikely. Including blas3 in the build increases
> >the speed of the code by a factor of 2-3. The jobs take either days or
> >weeks.
> > 
> >
> 
> Do you mean Atlas increases speed by 2-3x?  Or simply the BLAS3 bits 
> (non-Atlas)?  There are some BLAS3 libs running around in various 
> distributions.  SuSE will happily install BLAS for you if you request 
> it, in which case you need to worry about link order.
> 
> I would look at the output of ldd.  Make sure it is linking in the right 
> code.  Next, while the code is running, do an
> 
>    ps -fjg
> 
> after starting he code in another window.  You should see stuff like 
> this ...
> 
> crunch:/var/log/httpd # ps fjg
> PPID   PID  PGID   SID TTY      TPGID STAT   UID   TIME COMMAND
> 25381 25384 25384 25384 pts/7    27532 S        0   0:00 -bash
> 25384 27532 27532 25384 pts/7    27532 S        0   0:00  \_ tail -f 
> mail auth
> 25242 25245 25245 25245 pts/4    29077 S        0   0:00 -bash
> 25245 29077 29077 25245 pts/4    29077 R        0   0:00  \_ ps fjg
> 24959 24962 24962 24962 pts/2    28718 S        0   0:00 -bash
> 24962 27588 27588 24962 pts/2    28718 S        0   0:04  \_ nedit 
> login.html
> 24962 27618 27618 24962 pts/2    28718 S        0   0:00  \_ nedit 
> ../etc/bbs.ini
> 16116 16118 16118 16118 pts/1    23153 S        0   0:00 bash
> 16118 23153 23153 16118 pts/1    23153 S        0   5:26  \_ ping indy
> 
> Your threads will have different process IDs.
> 
> also try an
> 
>    ps fjma
> 
> and look for your process threads.
> 
> Joe
> 
> >Art
> >
> >On Tue, Aug 17, 2004 at 01:05:13PM -0400, Joe Landman wrote:
> > 
> >
> >>Art:
> >>
> >>Is it possible that the parallel section completes very quickly?  Or 
> >>that code path is not taken?
> >>
> >>Joe
> >>
> >>Art Edwards wrote:
> >>
> >>   
> >>
> >>>I mean that when we run the threaded code and look at top, don't see two
> >>>processors running. The load is always less than or equal to one. When
> >>>we run two separate jobs, we see a load of 2. 
> >>>
> >>>Art Edwards
> >>>
> >>>On Tue, Aug 17, 2004 at 07:45:32AM -0700, Alvin Oga wrote:
> >>>
> >>>
> >>>     
> >>>
> >>>>hi ya
> >>>>
> >>>>yup.. that is dual cpu
> >>>>
> >>>>On Tue, 17 Aug 2004, Art Edwards wrote:
> >>>>
> >>>> 
> >>>>
> >>>>       
> >>>>
> >>>>>We are running SuSE 9.1 on a system with 8 GB of ram and two scsi 
> >>>>>disks.
> >>>>>Here is /proc/cpuinfo
> >>>>>processor       : 0
> >>>>>   
> >>>>>
> >>>>>         
> >>>>>
> >>>>good
> >>>>
> >>>> 
> >>>>
> >>>>       
> >>>>
> >>>>>processor       : 1
> >>>>>   
> >>>>>
> >>>>>         
> >>>>>
> >>>>good
> >>>>
> >>>> 
> >>>>
> >>>>       
> >>>>
> >>>>>Linux bladerunner 2.6.4-52-smp #1 SMP Wed Apr 7 01:58:54 UTC 2004 
> >>>>>x86_64
> >>>>>x86_64 x86_64 GNU/Linux
> >>>>>   
> >>>>>
> >>>>>         
> >>>>>
> >>>>good
> >>>>
> >>>>
> >>>> 
> >>>>
> >>>>       
> >>>>
> >>>>>>Newbie question:
> >>>>>>
> >>>>>>We just purchased a dual opteron box that is running an SMP kernel. We
> >>>>>>have built a fortran code using g77 and a pthreaded atlas. Currently, 
> >>>>>>we
> >>>>>>see only one processor. What really simple thing have we missed?
> >>>>>>     
> >>>>>>
> >>>>>>           
> >>>>>>
> >>>>so what do you mean by "see only one processor" ??
> >>>>
> >>>>c ya
> >>>>alvin
> >>>> 
> >>>>
> >>>>       
> >>>>
> >>>
> >>>     
> >>>
> >>-- 
> >>Joseph Landman, Ph.D
> >>Founder and CEO
> >>Scalable Informatics LLC,
> >>email: landman at scalableinformatics.com
> >>web  : http://www.scalableinformatics.com
> >>phone: +1 734 612 4615
> >>   
> >>
> >
> > 
> >
> 
> 
> -- 
> Joseph Landman, Ph.D
> Founder and CEO
> Scalable Informatics LLC,
> email: landman at scalableinformatics.com
> web  : http://www.scalableinformatics.com
> phone: +1 734 612 4615

-- 
Art Edwards
Senior Research Physicist
Air Force Research Laboratory
Electronics Foundations Branch
KAFB, New Mexico

(505) 853-6042 (v)
(505) 846-2290 (f)
_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list