Having kernel networking problems with 681 kernel???

Stephen J. Smoogen smooge at gmail.com
Sat Dec 4 18:59:57 UTC 2004


I am trying to debug some new FC3 boxes and I am trying to figure out
some issues that seem to be kernel related, but I could be barking up
the wrong tree.

Our current pcap collectors and NFS servers are Debian woody boxes
with a mixture of unstable patches on them. We had some hard-drive
failures, and I snapped on Fedora Core 3 with updates on those to a)
because I had a hard time with Debian installer b) I wanted to get the
systems for RHEL-4 when it comes out (to meet some security
requirements we are being handed).

The way the systems are set up at the moment are something like this

[pcap collector] -------- [data analyzer]
 Debian                       FC3

All the network cards are E1000's and the link between the 2 boxes is
a cross-over optic fiber link. The pcap collector grabs its data and
then FTP's it over to the data analyzer every couple of minutes (the
files are 6 to 15 gigabytes in size and NFS was too slow/borked when
this was originally designed in 1999). If the connection goes down,
the Pcap box spools them until the Analyzer is ready. The problem we
are seeing is that when transferring multiple large and small files
over to the Analyzer the FC3 box seems to get into dazed and confused
mode. The vsftpd server accepts files at 30mb/s for a bit, then either
stops all together or slows down to 200 kb/s and doesnt pick up.

I have played with every option I can in the vsftpd configuration and
cant find anything that makes a difference. The only issue I can find
is that various kernel 'daemons' go into D state when this occurs
(pdflush, kswapd, kjournald).

A second 'related' issue is that I am trying to explain to the users
why ping is so much slower than with Debian. The Debian NetKit ping
will do a flood ping of 10000 1500sized packets in 30 seconds. The FC3
takes 3 minutes with what looks like some sort of backoff algorithm to
play nicely. I am not sure if it is truely related, but the fellows
who I am working with have been doing networking since 84 and it is
not what they are wanting.

Any help would be appreciated.

Stephen


-- 
Stephen J Smoogen.
CSIRT/Linux System Administrator




More information about the test mailing list