As an extra data point my colleague and I were looking at these bench
marks, and he commented that he'd seen other Linux graphs that looked
similar.  His speculation was that the irratic behavior could be caused by
caching effects or something like that.  In any case, the consensus, around
here anyway, is that at high loads, Linux just behaves like that, and
performance is just inconsistent.
-Brian

On Sep 28,  8:39pm, Andrew Doran wrote:
} Subject: Re: Thread benchmarks
} On Fri, Sep 28, 2007 at 02:27:34PM -0400, Thor Lancelot Simon wrote:
} 
} > On Fri, Sep 28, 2007 at 06:50:16PM +0100, Andrew Doran wrote:
} > > 
} > > And NetBSD-current compared to other systems:
} > > 
} > >         http://www.netbsd.org/~ad/sysbench/netbsd-and-others.png
} > 
} > Something interesting's happening in the Linux line on the graph right
} > at the right edge of the plotted region (20 threads).  Could you perhaps
} > run NetBSD-current against Linux again with the maximum number of threads
} > ramping up to 40, to see what the two curves look like as we head in
} > that direction?
} 
} I have also tried 10-100 and 100-1000 client connections. I don't have the
} numbers at hand, but Linux peaks around 550 tps somewhere around 100 client
} connections. The numbers I was getting from Linux were quite erratic and I
} had to throw out a few sets of results where the downward spikes were so bad
} that the results were basically useless.
} 
} > Either we degrade a lot more gracefully than Linux under load, or there's
} > an artifact in the Linux graph.  The current plot makes it impossible to
} > tell which, though.
} 
} In the long run Linux will beat NetBSD. That said it the behaviour I saw on
} this test cannot be called graceful!
} 
} Thanks,
} Andrew
quoted text
>-- End of excerpt from Andrew Doran