Date: Fri, 08 Oct 2010 09:09:35 -0500 From: Tom Judge <tom@tomjudge.com> To: a.smith@ukgrid.net Cc: freebsd-net@freebsd.org Subject: Re: bge watchdog timeout errors FreeBSD 7.3 Message-ID: <4CAF261F.4090509@tomjudge.com> In-Reply-To: <20101008144851.12024aez0aoyd2qs@webmail2.ukgrid.net> References: <20100924021115.GI15014@michelle.cdnetworks.com> <20100924123938.80702gxrzyfpury0@webmail2.ukgrid.net> <20100924165452.GA19036@michelle.cdnetworks.com> <20100927122713.12822br1odth4sro@webmail2.ukgrid.net> <20100927165129.GA1435@michelle.cdnetworks.com> <20100928132445.72052m5iw9f41ns4@webmail2.ukgrid.net> <20100928214438.GC1252@michelle.cdnetworks.com> <20100928230824.10533d7gmmba360w@webmail2.ukgrid.net> <20100928222532.GD1252@michelle.cdnetworks.com> <20101006174508.16931ioisevmzhy8@webmail2.ukgrid.net> <20101006174139.GC14879@michelle.cdnetworks.com> <20101008134251.367570jyw4igvkg8@webmail2.ukgrid.net> <4CAF1E0E.1080608@tomjudge.com> <20101008144851.12024aez0aoyd2qs@webmail2.ukgrid.net>
next in thread | previous in thread | raw e-mail | index | archive | help
On 10/08/2010 08:48 AM, a.smith@ukgrid.net wrote: > Quoting Tom Judge <tom@tomjudge.com>: > >> >> The R610 has a NetXtream II chipset which uses the bce(4) driver. If >> you are going to doing heavy network traffic, or using jumbo frames I >> would personally recommend you fit an Intel PCIe card to this server >> before you put it in production. >> > > Ah, ok thanks for the feedback. I dont have any intention of using > jumbo frames, but I expect it to work on a gigabit switch port. Do you > have personal experience that these or bad, or what is the reason not > to use this chipset? > The hardware its self seems fine, but there are some outstanding driver issues: 1) Jumbo frames will cause mbuf zone memory fragmentation leaving to dropped frames/stack lock ups (visual indication). There is a work around for this issue (define BCE_HDR_SPLIT in sys/dev/bce/if_bcereg.h). 2) There are performance issues with streams of small frames (up to about 200 bytes) where the RX ring fills up and the firmware starts dropping packets. Flow control will /fix/ the issue but really it just plasters over the issue and moves the frame drops to the egress queue on the switch. These are the current issues and they affect all Dell 10th and 11th gen hardware (2950,1950,R710,R610,R410,etc,etc). There some patches floating around that attempt to address these issues. * One to turn header splitting into a loader tunable. * Flow control is supported on all stable branches and head. I don't think it is in a release. * There is a patch to be able to tune the size of the ring buffers used by the card but from my testing it is unstable, and unusable with anything but the default values. I have spent a week trying to get the nic's stable in a R710 after which management pulled me off the work and authorized funds to fit Intel cards to the boxes. Tom -- TJU13-ARIN
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?4CAF261F.4090509>