Re: bash performance
----- "Martin" <inkubus@interalpha.co.uk> wrote:
> On Mon, 2010-10-25 at 23:31 +0300, Tarko Tikan wrote:
> > hey,
> >
> > This is a fresh squeeze install on T5220 (same for the laptop I'm
> comparing with). Whole machine seems strangely slow but one thing I
> can point out for sure is bash. I've been debugging this issue for a
> week now (even comparing strace -ttt time outputs with random intel
> machine). Does this seem strange to anyone else?:
> >
> > randomx86laptop$ time (VAR=0; for i in `seq 0 99999`; do let
> VAR=VAR+i; done)
> > real 0m1.235s
> > user 0m1.212s
> > sys 0m0.036s
> >
> > niagara2$ time (VAR=0; for i in `seq 0 99999`; do let VAR=VAR+i;
> done)
> > real 0m11.193s
> > user 0m10.777s
> > sys 0m0.388s
> >
> > I first noticed this with bash_completion - every new bash instance
> takes 2-3 seconds to start.
> >
> > Anyone have an idea whats going on?
>
> Can you give us a bit more info about the laptop? IIRC a T5220 runs
> an
> UltraSPARC T2 at about 1Ghz. These are designed for throughput, not
> latency. The T2 is out of order but only 2 way dispatch (IIRC),
> compared to a 3.4Ghz Intel chip with 4 way dispatch and full out of
> order instruction completion. Thus in terms of single threaded
> performance, I'm not exactly surprised an x86 laptop is 'faster'. A
> factor of 10 seems a bit steep but 3-4 maybe 5 times I could believe.
> However if you look at performance of code that is tightly branching,
> memory latency/bandwidth sensitive or thread interaction critical,
> you
> should see scaling up to .. well ... it depends on the app but I've
> seen
> near linear speed up, up to 32 threads on T1s.
>
> What other benchmarks have you tried? Also have you tried with any
> other OS (i.e. Open Solaris) on the same hardware; this should rule
> out
> anything Debian / Linux specific. Likewise have you tried other
> machines?
>
> Cheers,
> - Martin
>
>
Just for comparison purposes, the same command gives me the following on a much older Sunfire V100 (Ultrasparc IIe @ 500mhz) running Debian, and another of the same machine running Solaris 9:
Debian 2.6.18-6-sparc64:
real 0m10.534s
user 0m10.213s
sys 0m0.268s
Solaris 9:
real 0m10.985s
user 0m9.620s
sys 0m1.010s
And then slightly faster on a Sun T6320 running Solaris_11_b130 (Ultrasparc T2 @ 1.2ghz)
real 0m8.548s
user 0m8.564s
sys 0m0.045s
-Brian
Reply to: