XANSYS Message: 283 [Go back to message list]
[bookmark on del.icio.us]
No rating yet
Rate item:

Subject: Re: Multiple CPUs in SUN environment
Author: John Swanson
Date: 1999-01-16 10:31:00

John's Second Law of Computing would seem to apply here. The Second
Law is:

"If you have more users (tasks) than you have processors, do not
run in parallel."

The first corollary is:

"If you have more processors than users, you should run
in parallel."

In other words, on a lightly loaded machine (of which you don't
seem to have any), you
should request multiple CPU's, but in a heavy environment you are
better off using the machines
as thru-put machines. The best parallel performance is separate jobs,
rather than trying to
achieve parallelism within each job. In general, the higher the level
of parallelism, the more
efficient the process, and job level parallelism is about as high as you
can get.
John

> -----Original Message-----
> From: Dan Bohlen [SMTP:dan.bohlen@a...]
> Sent: Saturday, January 16, 1999 8:53 AM
> To: xansys@o...
> Subject: [xansys] Re: Multiple CPU's in SUN environment

> From: Dan Bohlen

> We also use SUN servers as our main compute boxes. I agree with Joe
> that the cpu time is the total ( take into consideration J Swanson's
> comments! ). What we have found is that for the same job with
> multiple
> cpus the cpu TOTAL will be higher than single threaded. We've assumed
> this is accounting for the cpu "overhead" of running in parallel. Our
> wall clock times (relative to the cpu time for THAT run not compared
> to
> the 1 cpu case.) are 10-30% less than the cpu seconds (with 3 cpus
> turned on). I show a type run below where we would count that as a 22%
> savings. (But what would the cpu seconds have been single threaded We
> have also found that there are times when the "overhead" can become a
> significant % of the "baseline" single threaded cpu time, thus
> increasing not only the cpu time but also the wall clock versus a 1
> cpu
> run. We seen this with smaller jobs generally (10 minutes).

> What gets debated here (and I'd appreciate an insights from the list)
> is
> whether we should be running our big jobs single or multiple threaded.
> For a lightly loaded machine the answer seems pretty straight forward

> multiple cpus (>1cpu). But our SUN boxes generally are pretty loaded
> and there is a camp here that theorizes that with 9 >1cpu jobs that
> overhead just piles up with cpu swapping and you actually slow all the
> jobs down relative to running them single threaded. Who's right? Is
> the single cpu job cpu time dependant on system load too? And the
> elapsed time, I assume that is also a function of I/O as well as cpu
> so
> system load again comes into play. One piece of data I do have from
> our
> SUN systems folks here - our SUNs are almost never cpu limited but I/O
> limited! Probably partly due to the virtual memory configuration
> mentioned in another part of this discussion.

> Of course we could run tests to get some data, but taking a SUN
> off-line
> would cause a minor revolution here. Also sad to report that this
> type
> of testing didn't happen before we bought the things. So I'd
> appreciate
> any tips. TIA

> -----------TYPICAL 3 cpu output --------

> | ANSYS RUN
> COMPLETED |

> | RELEASE 5.4 SP UP19980310
> SOLARIS |

> | CP TIME (sec) = 3822.364 TIME =
> 10:46:25 |
> | ELAPSED TIME (sec) = 2968.000 DATE =
> 01/14/1999 |

> more....

> b1250.out: | CP TIME (sec) = 8273.009 TIME =
> 19:26:44 |
> b1250.out: | ELAPSED TIME (sec) = 8543.000 DATE =
> 01/12/1999 |
> b184.out: | CP TIME (sec) = 5103.140 TIME =
> 12:44:37 |
> b184.out: | ELAPSED TIME (sec) = 6426.000 DATE =
> 01/13/1999 |
> b351.out: | CP TIME (sec) = 10689.620 TIME =
> 12:49:39 |
> b351.out: | ELAPSED TIME (sec) = 11796.000 DATE =
> 01/12/1999 |
> b7400.out: | CP TIME (sec) = 10495.750 TIME =
> 14:45:42 |
> b7400.out: | ELAPSED TIME (sec) = 13624.000 DATE =
> 01/14/1999 |
> b7471.out: | CP TIME (sec) = 4894.610 TIME =
> 12:25:40 |
> b7471.out: | ELAPSED TIME (sec) = 5314.000 DATE =
> 01/13/1999 |

> "Metrisin, Joseph T." wrote:

> > From: "Metrisin, Joseph T."

> > Geno,

> > If you are looking at the CPU time printed at the bottom of the
> > ANSYS output, this is the TOTAL CPU time accumulated by
> > all four processors. In other words, your job still takes nearly
> > the same amout of CPU time to solve, it's just spread out over
> > the various processors. If you look at the elapsed time, it
> > will probably show that it solved in roughly 1/4 of the wall clock
> > time.

> > Joseph T. Metrisin

> > Structural Methods
> > Pratt & Whitney/Unitetd Technologies
> > M/S: 714-03
> > P.O. Box 109600
> > West Palm Beach, FL. 33410-9600

> > Phone: (561) 796-5967
> > Fax: (561) 796-8993

> > > From: Geno Vallejos[SMTP:gvalle@c...]
> > > Sent: Friday, January 15, 1999 12:01 PM
> > > To: xansys@o...
> > > Subject: [xansys] Multiple CPU's in SUN environment

> > > From: Geno Vallejos

> > > To all my fellow ANSYS users:

> > > I have had some interesting results when solving the same
> structural
> > > problem using the same model but on two different SUN machines.
> > > Below are the specifics;

> > > The structural model is a part created in ProENGINEER 18
> > > I am using ANSYS ProFEA 5.4 and ProENGINEER 18.

> > > Solution #1:
> > > Machine: SUN E-3000 server
> > > Processors: 4 each 360 MHz CPU's
> > > Operating System: Solaris 2.6
> > > CP Time (from the file.lst summary) = 10 minutes 28 seconds

> > > The first time I requested four CPU's using the config.ans file,
> with the
> > > following line added;
> > > NUM_PROC = 4
> > > The config.ans file was located in the /docu directory.
> > > The file.lst indicated that 4 CPU's will be used

> > > The second time I requested four CPU's using a .inp file. The
> .inp file
> > > had
> > > the following line;
> > > /config,nproc,4

> > > The solution times (CP Time listed above) was 10 minutes 28
> seconds

> > > Solution #2:
> > > Machine: SUN Ultra 60 workstation
> > > Processors: 1 each 360 MHz CPU
> > > Operating System: Solaris 2.6
> > > CP Time (from the file.lst summary) = 10 minutes 53 seconds

> > > Can anyone help explain why the E-3000 server and the Ultra 60
> > > workstation solved the structural analysis problem by taking the
> same
> > > amount of time????

> > > I have heard some ideas....but nothing concrete. Some here at ATL
> say
> > > that the four SUN processors will work as hard as they need to to
> solve
> > > the problem. I say that I would like them to solve the problem as
> fast as
> > > possible. Observing the CPU activity bar during the solution
> phase
> > > indicated that at no time were any of the four CPU's working any
> higher
> > > than 40-50% of maximum capacity. Was this due to some SUN or
> Solaris
> > > setting "chocking" the CPU speed or efficiency; or some ANSYS
> setting
> > > not getting the full support of multi-tasking from the SUN
> server??

> > > Thanks in advance!

> > > Geno Vallejos
> > > ATL Ultrasound - A Philips Company
> > > Bothell, Washington 98021
> > > www.atl.com

> > > to digest, go to the ONElist web site, at
> and

> Dan Bohlen Phone: (513) 243-4645 / Dial Comm [8*332-4645]
> Stress Analyst Email: dan.bohlen@a... M/D K-105
> GE Aircraft Engines Fax: (513) 243-4888 or 243-4886

Posts possibly associated with message #283AuthorDateScore
247Multiple CPUs in SUN environmentGeno Vallejos gvalle@1999/01/15 
248Re: Multiple CPUs in SUN environmentBrian Breaux1999/01/15 
250Re: Multiple CPUs in SUN environmentJoe Metrisin1999/01/15 
260Re: Multiple CPUs in SUN environmentBrian Murphy1999/01/15 
266Re: Multiple CPUs in SUN environmentJohn Swanson1999/01/15 
269Re: Multiple CPUs in SUN environmentPatrick Tibbits1999/01/15 
270Re: Multiple CPUs in SUN environmentHeinz.Hempowitz@1999/01/15 
281Re: Multiple CPUs in SUN environmentDan Bohlen1999/01/16 
283Re: Multiple CPUs in SUN environmentJohn Swanson1999/01/16 
323Summary of Multiple CPUs in SUN environment questionGeno Vallejos gvalle@1999/01/19 
324Re: Summary of Multiple CPUs in SUN environment queChristopher Wright1970/01/01 
332Re: Summary of Multiple CPUs in SUN environment questionJoe Rasche1999/01/20 
340Re: Summary of Multiple CPUs in SUN environment questionBrian Murphy1999/01/20