Date: Fri, 29 Oct 2010 10:23:19 -0700 From: Chuck Swiger <cswiger@mac.com> To: =?utf-8?B?0JrQvtC90YzQutC+0LIg0JXQstCz0LXQvdC40Lk=?= <kes-kes@yandex.ru> Cc: freebsd-net@freebsd.org Subject: Re: Re[2]: Polling slows down bandwidth Message-ID: <FA2E3C9B-A1A1-4FB3-80AD-26F5518C1ABF@mac.com> In-Reply-To: <606859717.20101029093926@yandex.ru> References: <1519248747.20101028232111@yandex.ru> <1452146D-A590-4676-A662-14D0EEE82152@mac.com> <606859717.20101029093926@yandex.ru>
next in thread | previous in thread | raw e-mail | index | archive | help
On Oct 28, 2010, at 11:39 PM, =EB=CF=CE=D8=CB=CF=D7 =E5=D7=C7=C5=CE=C9=CA =
wrote:
> =FA=C4=D2=C1=D7=D3=D4=D7=D5=CA=D4=C5, Chuck.
Um, greetings? =20
> =F7=D9 =D0=C9=D3=C1=CC=C9 28 =CF=CB=D4=D1=C2=D2=D1 2010 =C7., =
23:41:58:
>=20
> CS> On Oct 28, 2010, at 1:21 PM, =EB=CF=CE=D8=CB=CF=D7 =E5=D7=C7=C5=CE=C9=
=CA wrote:
>>> [ ... ]
>=20
> CS> What is "sysctl kern.clockrate", and have you increased kern.hz
> CS> in /boot/loader.conf to at least 1000, if not 2000 or 4000?
>=20
> # vmstat -i
> interrupt total rate
> irq14: ata0 193948 6
> irq16: rl0 42829515 1464
> irq23: nfe0 41224044 1409
> cpu0: timer 58494158 1999
> irq256: igb0 106911 3
> irq257: igb0 254606 8
> irq258: igb0 2 0
> Total 143103184 4892
>=20
> # sysctl kern.clockrate
> kern.clockrate: { hz =3D 1000, tick =3D 1000, profhz =3D 2000, stathz =
=3D 133 }
>=20
> # sysctl kern.hz
> kern.hz: 1000
> but I have configured and installed kern with 2000HZ
> "systat -v" shows that: 2002 cpu0: time
Actually, the interrupt rate is tracking profile hz, which is roughly =
double the actual kern.hz-- per sysctl, you should try to at least =
double kern.hz.
>=20
> CS> Polling mode operation generally performs better when using older
> CS> 100Mbs ethernet NICs which do not support interrupt mitigation and
> CS> various capabilities like TSO4; gigabit ethernet NICs are smarter
> CS> hardware and can generally outperform polling mode.
>=20
> so using polling on gigabit NICs is a bottle neck? and is cause of low =
performance, is not?
Simple answer is yes. It should be possible that you could tune polling =
to get similar performance, or at least better performance than you see =
now, but the additional hardware capabilities of gigabit NICs are likely =
to outperform polling mode, just as polling mode can generally =
outperform old 100MBs ethernet NICs.
Regards,
--=20
-Chuck
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?FA2E3C9B-A1A1-4FB3-80AD-26F5518C1ABF>