[Click] click performance test
Robert Morris
rtm at amsterdam.lcs.mit.edu
Mon Sep 22 15:40:18 EDT 2003
85 cycles sounds similar to the value we measured for the cost of
rdtsc a few years ago. rdtsc waits for the CPU pipelines to flush, so
it's expensive.
> To: jblicn at yahoo.com.cn
> Subject: Re: [Click] click performance test
> Date: Mon, 22 Sep 2003 11:35:47 -0700
> From: Eddie Kohler <kohler at icir.org>
> cc: click at amsterdam.lcs.mit.edu
>
> Hi Jiangbo,
>
> > I am doing performance test on click elements. I use RDTSC
> > to read the performance counter. In your thesis, you got the
> > very accurate results like checkipheader takes 457ns/packet.
> > How can you get the result so accurate?
>
> As I remember, we report the mean of several trials. We didn't report
> standard deviation numbers, so I couldn't tell you the error range. As the
> thesis says,
>
> Each element's cost is the difference between the Pentium III cycle counter
> value before and after executing the element, decreased by the time
> required to read the cycle counter; it includes the virtual function
> call(s) that move a packet from one element to the next.
>
> > Another problem is when I do two continuous RDTSC, It took 85
> > cycles. Seems to much. Is there anything I missed?
>
> Probably not; rdtsc is quite expensive; but I'm not positive about
> particular numbers.
>
> Eddie
> _______________________________________________
> click mailing list
> click at amsterdam.lcs.mit.edu
> https://amsterdam.lcs.mit.edu/mailman/listinfo/click
More information about the click
mailing list