Tony Nelson wrote: > At 10:44 PM -0600 11/14/07, blake@xxxxxxxx wrote: > >> I seem to be having trouble sending mail from an FC5 server to >> mail.jococourthouse.com. >> >> The problem manifests itself as extreme slowness, which eventually leads >> to a timeout before I can even receive a full welcome prompt from the >> remote side. >> >> I have tested on ~ 10 boxes (different hardware) around the US ranging >> > >from FC5 through FC8 and NONE can telnet to this mx while RH, FC3 RH4, > >> windows, and cisco gear on the same networks can ALL telnet without a >> problem. >> >> wireshark on the FC5 boxes show that the TCP handshake occurs as normal, >> however from then on all I receive from the remote side are 60 byte frames >> labeled 'TCP segment of a reassembled PDU' until I eventually receive an >> RST from the remote. - A capture on the working boxes looks normal. >> >> I suspect the problem is related to an IP/TCP change in the kernel or >> FC5+, however I have no idea what change could have occurred. Any insight >> or recommendations for resolving this problem would be appreciated. >> > > TCP Window Scaling, introduced about that time, and a bad router? > > > At 5:57 PM +0200 7/24/06, Alexander Dalloz wrote: > >> Stefan Kuhlemann schrieb: >> >> >>> even more strange - today around 10:00 o'clock the sendmail started to >>> work normal until 16:00 o'clock. (with nothing changed since sunday >>> 22:00) >>> >>> My feelings tell me that the reason might be some strange 'ill-broken' >>> firewall network stack or some weird netwok communication issues (like >>> the 'fun' with the ECN-flag some years ago) >>> I'm tomorrow 'on-site'...... >>> >>> Stefan >>> >> Try following kernel parameter change >> >> sysctl -w net.ipv4.tcp_window_scaling=0 >> >> and see if that helps. TCP window scaling behaviour changed with kernel >> 2.6.17 to be again more aggressive[1]. Some other systems are just >> broken in this regard. >> >> Alexander >> Tony, disabling window scaling allows me to connect to the remote smtp server. I had thought this might be the issue, however the FC3 boxes have this feature enabled and don't seem to have any problem. I thought there might be some happy medium (perhaps buffer sizes or scaling behavior) that might allow me to keep scaling enabled, but be less aggressive. I've been over the man pages for TCP and sysctl and I've tried a few settings with regard to rmem, wmem, etc but no luck. However, I did find a fix that is acceptible. You can set the scaling per host/network using the route command. http://slaptijack.com/system-administration/tcp-performance-tuning-and-broken-routers/ http://lwn.net/Articles/213987/ Thanks all that responded, -B
