+> On Tue, Oct 26, 1999 at 08:38:43AM -0400, Jasjeet Bagla wrote:
+> >
+> > Hi,
+> >
+> > I have an SMP box (PII 333MHz) with SuSE 6.1 running 2.2.13. I am using
+> > it to run large simulations. If i run one job, one cycle takes about 73
+> > minutes of CPU time. It i run two identical jobs, this goes up to a
+> > little more than 90 minutes. Both the jobs are big - they require 160MB
+> > of RAM each. I have 384MB so that should not be a problem, in principle.
+> > I suspect this is a scheduler problem. I want to know if something can
+> > be done about it.
+> >
+>
+> Since your jobs use a lot of memory, I suspect that they also use
+> the memory a lot. That is, they perform operations on a multi-megabyte
+> dataset repeatedly, so that the data set each CPU works on doesn't
+> fit well in L1/L2 cache and actual main memory access is required.
+>
+> If that holds, I guess what you're seeing is memory contention.
+>
+> Memory access has been improved greatly on newer boards, but on my
+> dual PPro box at home, I can get ~100% speedup when the problems fit
+> in cache, and roughly 10-20% speedup on memory intensive problems.
+>
+> You still see pretty good speedup (90 vs. 146 minutes).
+>
+> By the way, you're not having anything else running are you ?
+> Things like X, an editor, etc. will gladly run on the spare CPU while
+> your simulation runs on one CPU. But with two simulations you will
+> actually see the performance hit.
Screensavers are an exceptionally "expensive" way of wasting CPU cycles...
Just my 0.02 $
Herbert
==============================================================================
If god would have intended man to smoke, he would have set him on fire.
-
Linux SMP list: FIRST see FAQ at http://www.irisa.fr/prive/dmentre/smp-howto/
To Unsubscribe: send "unsubscribe linux-smp" to [EMAIL PROTECTED]
