Yeah, that did the job (at least the part with the packets larger than MTU), I still have package disordering but now I can take this to my ISP. Thanks. On Wednesday 26 March 2008 18:37:52 Bill Fink wrote: > On Wed, 26 Mar 2008, Pavel Georgiev wrote: > > I have several linux boxes hooked to a gigabit switch and accidentally > > noticed that tcpdump reports packages with size larger than 1500: > > > > tcpdump greater 1600 -n > > > > This produces many lines of output with packets with size > 1500 (2960, > > 4420, etc.) > > > > The MTU of the interface is 1500 (reported by ifconfig), the MTU on the > > other boxes hooked to the same switch is also 1500. > > > > This would be OK if it was limited to the intranet only (although strange > > as the MTU is 1500 on all boxes), but this also happens for packets going > > out to the Internet. I traced a tcp connection to an outside host on both > > sides and notices that both hosts send an mss of 1460 during connection > > initialization, but the first packet (and all other) that the box in > > question sends is with length 2960, which on the other side is received > > and acknowledged as two packets with length 1500. This suggest that > > someone on the way (probably the first router) does fragmentation, which > > is OK. The problem is that eventually packets on the remote host start to > > arrive with changed order, which tcp handles as expected, but this > > somehow prevents the tcp window from growing and as a result the transfer > > rate never gets high enough to utilize the line. > > > > What I`m trying to find out is how come packets leaving my box have > > larger size than the MTU of interface (I`m guessing that if I solve that, > > no additional fragmentation will occur, packets will arrive in the right > > order and tcp windows will grow as expected). > > > > Few notes: > > - kernel on the boxes in question: 2.6.14.6 and 2.6.15 > > - the line to the boxes is wide enough to handle 100Mbit/s, if I try the > > same test that I describe but in the other direction, I get the full > > 100Mbit/s speed as tcp window grows high enough to allow this even with > > latency > 100ms. > > - boxes in question use e1000 driver. > > Perhaps try disabling TSO on the interface with ethtool. > > -Bill > -- > To unsubscribe from this list: send the line "unsubscribe linux-net" in > the body of a message to majordomo@xxxxxxxxxxxxxxx > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-net" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
