Date: Tue, 20 Nov 2012 13:52:08 -0800 (PST) From: Barney Cordoba <barney_cordoba@yahoo.com> To: khatfield@socllc.net Cc: freebsd-net@freebsd.org Subject: Re: FreeBSD boxes as a 'router'... Message-ID: <1353448328.76219.YahooMailClassic@web121602.mail.ne1.yahoo.com> In-Reply-To: <1281530059.17550.1353434227450@238ae4dab3b4454b88aea4d9f7c372c1.nuevasync.com>
next in thread | previous in thread | raw e-mail | index | archive | help
Anyone who even mentions polling should be discounted altogether. Polling= =0Ahad value when you couldn't control the interrupt delays; but interrupt= =0Amoderation allows you to pace the interrupts any way you like without=0A= the inefficiencies of polling.=0A=0AThe idea that polling uses less CPU is = complete baloney; its exactly=0Athe same code as without polling. It's some= thing else that the CPU=0Ahas to do, so polling is a negative with almost a= ny modern NIC. With=0Aintel nics you can set whatever interrupt rate you wa= nt; setting your=0Aints/sec to whatever hz is (or whatever polling timer is= used) has=0Athe same effect without having to have hacked code running in = your driver.=0A=0A--- On Tue, 11/20/12, khatfield@socllc.net <khatfield@soc= llc.net> wrote:=0A=0A> From: khatfield@socllc.net <khatfield@socllc.net>=0A= > Subject: Re: FreeBSD boxes as a 'router'...=0A> To: "Victor Balada Diaz" = <victor@bsdes.net>=0A> Cc: "freebsd-isp@freebsd.org" <freebsd-isp@freebsd.o= rg>, "John Fretby" <jfretby@googlemail.com>=0A> Date: Tuesday, November 20,= 2012, 12:57 PM=0A> One thing I have noticed is mixed=0A> results with poll= ing depending on the version.=0A> =0A> My experience with similar NICs was = that polling increased=0A> the PPS capabilities up to 7.4 but post 7.4 we h= ave seen=0A> most cases where polling caused either connectivity issues=0A>= or decreased overall performance.=0A> =0A> Now we were running full 1Gbps = in our tests. With only=0A> 140Mbps you should be able to handle this amoun= t without=0A> polling or additional kernel tweaks. Specifically with 9 - I= =0A> would recommend doing needed sysctl tweaks without polling=0A> and as = long as you are not receiving DDoS traffic then it=0A> should prove perfect= ly stable.=0A> =0A> =0A> =0A> On Nov 20, 2012, at 11:46 AM, "Victor Balada = Diaz" <victor@bsdes.net>=0A> wrote:=0A> =0A> > On Tue, Nov 20, 2012 at 03:3= 5:13PM +0000, John Fretby=0A> wrote:=0A> >> Howdy all,=0A> >> =0A> >> We've= currently got an ageing HP DL360 running as a=0A> 'router' - it has=0A> >>= 100Mbit in/out onto our network, and has two 'bce'=0A> NIC's providing in/= out.=0A> >> It's running quite an old version of FreeBSD (6 I=0A> think) - = but works.=0A> >> =0A> >> As the network gets busier we've noticed the amou= nt=0A> of interrupt time on it=0A> >> is climbing (as you'd expect - i.e. e= sp. if many=0A> small packets are being=0A> >> forwarded). Many moons ago w= e did experiment with=0A> this box - and enabled=0A> >> device polling (inc= . upping the HZ on the box and=0A> recompiling the kernel=0A> >> etc). This= didn't work very well at the time=0A> (probably because it was in=0A> >> i= t's infancy) so we left it off in the end.=0A> >> =0A> >> If we were to rep= lace this box, with something new=0A> - say a SuperMicro based=0A> >> syste= m with two:=0A> >> =0A> >>=A0=A0=A0Intel 82574L's (em Driver Based)=0A> >> = =0A> >> And enable polling - is it likely to "just work"=0A> these days? Th= e current=0A> >> upstream is 100Mbit, we're looking to upgrade this=0A> to = 1Gbit in, but with=0A> >> say 200Mbit comitted on it (so shouldn't go above= =0A> 200Mbit).=0A> >> =0A> >> Is there anything that has to be done to enab= le=0A> polling - other than=0A> >> recompiling GENERIC to support it? - i.e= . no HZ=0A> hacks or anything needed on=0A> >> 'modern' machines (it's a qu= ad core Xeon).=0A> > =0A> > Hello John,=0A> > =0A> > You might find interes= ting to read this thread:=0A> > =0A> > http://lists.freebsd.org/pipermail/f= reebsd-current/2012-November/037590.html=0A=0A> > =0A> > In short: device p= olling can decrease performance on=0A> modern hardware.=0A> > =0A> > You mi= ght want to try upgrading to a new FreeBSD=0A> version and tuning it someho= w=0A> > before buying a new server. More info on tuning the=0A> network sta= ck:=0A> > =0A> > http://wiki.freebsd.org/NetworkPerformanceTuning=0A=0A>; > = =0A> > Regards.=0A> > Victor.=0A> > -- =0A> > La prueba m=E1s fehaciente de= que existe vida=0A> inteligente en otros=0A> > planetas, es que no han int= entado contactar con=0A> nosotros. =0A> > _________________________________= ______________=0A> > freebsd-isp@freebsd.org=0A> mailing list=0A> > http://= lists.freebsd.org/mailman/listinfo/freebsd-isp=0A=0A> > To unsubscribe, sen= d any mail to "freebsd-isp-unsubscribe@freebsd.org"=0A> ___________________= ____________________________=0A> freebsd-isp@freebsd.org=0A> mailing list= =0A> http://lists.freebsd.org/mailman/listinfo/freebsd-isp=0A>; To unsubscri= be, send any mail to "freebsd-isp-unsubscribe@freebsd.org"
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?1353448328.76219.YahooMailClassic>