North American Network Operators Group|
Date Prev | Date Next | Date Index | Thread Index | Author Index | Historical
2008.02.19 NANOG 42 JumboFrames across the Interet
*whew* Two more to go. :) Matt 2008.0-2.19 Tom Scholl, ATT labs, Jumbo packets on the internet MTU = maximum transmit unit PMTUD = path mtu discovery Every media has its own 'standard' MTU. Most have large MTU of at least 4470 Ethernet MTU is 1500, which makes the large size in core mostly useless. benefits of bigger packets==fewer packets per second, fewer lookups, fewer interrupts. on the host, does it really help? What about jumboframes? what's a jumbo frame? anything bigger than 1500; modern gear can do larger than 1500; driven by customers, not by a standard. Path MTU discovery fragmenting packets and reassembling is difficult PMTUD detects lower MTU, sends message to host asking it to readjust packet size. PMTUD is very easy to break. filtering ICMP kills it if not done right. If mismatched MTUs on a link, PMTU can't detect it. path mtu mismatch can't be communicated back if the far end of a link has an MTU too small, as upstream never gets the message. Interprovider jumboframes works great on point to point links; but what about point to multipoint links? Can't negotiate to a given router, so has to be set across the fabric/subnet. how about adaptive arp protocol; use existing arp to find MAC address, then use that to probe for MTU on path. good luck getting it implemented. hacking it in BGP might work, but would require everyone to run BGP. some exchanges have different vlans for jumboframe (NetNOD) but we already have so many vlans! what's a good target MTU? How about 8192 plus lots of header room and encapsulation room. anything bigger than 1500 makes sense. What about going all the way up to 65k? how do you specify the MTU on an interface? is it frame payload frame payload plus headers ...plus FCS ...plus 802.1q depends on vendors and cards you're running. Not all vendors can do 9k older cisco gear older juniper PICs enabling jumbo can be production impacting. may be unrealistic to rexpect jumboframes to reach all the way to home users for any time soon. Actions required have IEEE standardize on a new MTU value? Need a negotiation to discover neighbor MTU need less breakable replacement for PMTUd http://darkwing.uoregon.edu/~joe/jumbo-clean-gear.html Q: better PMTUd--could end host report back to sending host that the biggest sized fragment I got back was Y, don't send packets bigger than that please? Sounds reasonable. Q: Todd, Renesys. Lots of challenges, and a few benefits; does this mean we're stuck with 1500? Will larger MTUs actually happen? Can we at least support 1500 across the board? larger would give more room for encapsulation as we add more and more header foo into the mix. Q: Danny notes that options aren't being calculated correctly; is MD5 calculated before or after CRC, etc. Q: Kevin Oberman, ESnet, RE networks around the world are generally running jumboframes, the joint engineering task force, under department of defense came up with recommended value of 9000 bytes, and the RnD community has standardized on 9000 payload bytes to be usable from end to end; intra-AS is your own business, but if you want to talk to the Q: Darrel, Calren, when you have a boundry, you have a big cloud of 9000 byte MTU, your edge device will have to handle the fragmentation; be aware of the performance impacts that you may cause in having edge router have to process and deal with or discard. Q: Patrick--we're basically talking about distributing the fragmentation processing challenge out towards the edge; sounds like a reasonable idea to implement. Q: David Sinn notes that he's dealt with issues at the gigapop where setting the DF bit is actually nicer than making the edge router try to handle the fragementation. What about using rate limiters Q: RAS, why are we doing it at all? IPv6 doesn't support it at all; why not just stop doing fragmentation across the board--if you forget to set DF, you just blackhole yourself. OK, Break time. Survey is linked off nanog.org, go fill it out! In spite of late start first two days, we start at 9am tomorrow--wake up extra early.