Date: Wed, 13 Aug 2008 22:25:03 -0700 From: "Jack Vogel" <jfvogel@gmail.com> To: Paul <paul@gtcomm.net> Cc: "freebsd-net@freebsd.org" <freebsd-net@freebsd.org> Subject: Re: HEADS UP: E1000 networking changes in STABLE/7.1 RELEASE Message-ID: <2a41acea0808132225s6e77175fnf6780b249ed07058@mail.gmail.com> In-Reply-To: <48A3BC67.3050905@gtcomm.net> References: <2a41acea0808131502y39879a22u39c472bd0b810fc2@mail.gmail.com> <48A3A2DA.8030404@gtcomm.net> <2a41acea0808132135oe9ebc6bk9423ac19f2e9f77a@mail.gmail.com> <48A3BC67.3050905@gtcomm.net>
next in thread | previous in thread | raw e-mail | index | archive | help
On Wed, Aug 13, 2008 at 10:02 PM, Paul <paul@gtcomm.net> wrote: > The hardware spec sheet says it does support multi queue but only MSI, so I > would suspect it is disabled by default since that would (maybe) cause > performance to be less. > Would it take advantage of multiple cpus to turn on the multiple queues > even without MSIX? (i.e. having more than 1 taskq in freebsd per interface > for receives/sends) > It should still do per-vector masking so it couldn't be THAT bad with the > multi queues enabled but there's probably a good reason why they are > disabled by default eh? Would be > fun to test it out though :> :> > I think even if the driver / bsd split it up into multiple taskq's so it > would take advantage of SMP better that it would work, but that might > require some modifications outside of the ome of his work within a couple weeks. I have also been working on what performance tweaks I havedriver level (which I can think of > a lot) but I guess i"m hoping for an easier solution so > I don't have to go replacing all these 82571EB cards.. They work great, but > the cpus are running out.. Having 3 or 4 idle cpus while the other ones are > struggling due to network load and dropping packets is aggravating :) Will > the 82575/6 cards alleviate that problem or will I still be faced with the > same issue (due to os limitations)? > I plan on buying the dual port ET controller[82576] (which was supoosed to > be shipping last month) but I can't find one anywhere! argh.. The Linux team here implemented multiqueue based on 82571 3 or 4 years ago, they found that without MSIX it just caused more problems than it was worth and they turned it off again until recently. Linux is in a better position right now, Dave Miller has been hacking madly at their stack and they have pretty complete multiqueue support in the stack. One of my friends here at Intel was responsible for a lot of the work and code along the way. I know there's stack work going on, hopefully all drivers will benefit from that as it comes out. Also Kip Macy has been doing some work for a customer that is based on my igb code, he promised to have some patches back to me in a couple weeks, I'm sure it will be good stuff too. I do not know how hard 82576 is to get right now, it has a TON of potential that I have no where near tapped yet. Before the firedrill the last couple of days I had started on updating the ioatdma driver, which we need so that I can put DCA into the igb driver (Direct Cache Access), that should be a big performance enhancer when I can get it done. Jack
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?2a41acea0808132225s6e77175fnf6780b249ed07058>