Re: DNS Performance Numbers

On Fri, 17 Nov 2006, Marcelo Gardini do Amaral wrote:

I made some tests using 7.x with all the debugging disabled:

queries / s

Int bind (d_t) bind (e_t) nsd (1_s) nsd (2_s)
--- ---------- ---------- --------- ---------

bge 15439 14733 12910 10946
em 37655 34092 42411 41974
...

Try my fix for bge's interrupt handling in freebsd-net.

I tried but I didn't get any improvement.

I also have some latency and overhead optimizations for bge. I think
your test uses many threads to get parallelism so latency optimizations
wouldn't help much for it, but perhaps the problem is that the type
of interrupt moderation used by bge accidentally limits parallelism,
and then my changes might help by accidentally increasing parallelism
(they are only intended to help much for the non-parallel case). bge
hardware has fairly sophisticated interupt moderation, but poor tuning
of it results in it reducing to almost the equivalent of polling at a
rate of 6667 Hz or 2*6667 Hz depending on whether the tx and rx polls
accidentally get in sync. (Intentional use of polling would give even
worse latency unless you pessimize HZ too.) To reach em's number of
operations per second, bge would have to do operations in bursts of 7
or 3.5 operations per poll, and the details of the operations or the
polling might prevent this.

For a quick test of latency reduction, try initializing
sc->bge_rx_coal_ticks to 150 instead of 1.

Bruce
_______________________________________________
freebsd-performance@xxxxxxxxxxx mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-performance
To unsubscribe, send any mail to "freebsd-performance-unsubscribe@xxxxxxxxxxx"

Relevant Pages

  • Re: Technical superiority of polling loops
    ... > more a watchdog loop than tight polling using sleep. ... > if it consumes around 1000 clocks, ... Low CPU load at the expense of latency. ...
    (alt.lang.asm)
  • Re: [Hdaps-devel] Generic interface for accelerometers (AMS, HDAPS, ...)
    ... input infrastructure cause a noticable latency increase compared to ... there's a small issue with polling frequency. ... parking schemes *aren't* implemented entirely in the kernel? ...
    (Linux-Kernel)
  • Re: Technical superiority of polling loops
    ... this loop has _horrible_ latency. ... if it consumes around 1000 clocks, ... Unless latency is unimportant and you believe polling is easier ...
    (alt.lang.asm)
  • Re: polling(4) rocks!
    ... > added latency typically ends up being around 1ms. ... Latency is affected if in the case of using interrupts, ... than for polling. ... If you plot a latency versus pps graph, ...
    (freebsd-net)
  • Re: Re[2]: routing for 1000 users and 10Mbit internet.
    ... > BM> Try polling and see if the load reduces and the performance increases. ... The man page for polling does not list the bge driver as supported, ... the fxp driver _does_ support polling. ...
    (freebsd-questions)