Ethernet Bonding Performance under kernel 2.4.0
Many of your questions may have already been answered in earlier discussions or in the FAQ. The search results page will indicate current discussions as well as past list serves, articles, and papers.
Jonathan Earle jearle at nortelnetworks.comTue Jan 16 11:48:31 PST 2001
- Previous message: Shield or skilled?
- Next message: Problems with 3c905C under Sycld
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Hi all,
I've a system comprosed of two PIII machines, equipped with Znyx 346Q 4port
ethernet cards (tulip driver) which I'd like to connect together in a bonded
configuration. For various reasons, we require 2.4.0 kernels on our
machines - currently we are using 2.4.0-test9.
The setup is simple: each port on a 346Q in one machine is connected to the
corresponding port on the 346Q in the other machine via a crossover cable.
+-------+ +-------+
| |------| |
-----| Box A |------| Box B |-----
| |------| |
| |------| |
+-------+ +-------+
Problem #1
----------
Initally, after bootup, the performance of each of the four networks between
the two PCs is subpar. Transfer rates will vary from a few hundred KB/s to
perhaps a few MB/s, and the transfer time is appreciably long. This, on a
forced 100TX-FD link. After a few minutes however, things appear to settle
down, and I can achieve 11.2MB/s when transferring a large binary file via
ftp (rate as reported by ncftp). The de4x5 driver shows the same behaviour.
Problem #2
----------
I built the bonding driver, and using a copy of ifenslave.c which I found
for kernel 2.3.50, I was able to make a bonded channel. The trouble I found
is that the performance was not at all what I expected. Using the first eth
port, I achieved a throughput (FTP transfer of a large binary file) of 10.4
MB/s (11.2MB/s if set to full duplex). Using 2 ports, the performance
dropped to about 3.5MB/s. Adding a third port brings the throughput to
about 5.2MB/s and adding the fourth port only takes it up to 5.75MB/s.
The de4x5 driver shows the same drop in performance as the tulip driver.
Using the TEQL (trivial link equalizer) (instructions from the Adv-routing
howto) provides the same measurements exactly.
Thoughts?
Cheers!
Jon
- Previous message: Shield or skilled?
- Next message: Problems with 3c905C under Sycld
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
More information about the Beowulf mailing list
