3c905TX RXs 10 times slower than TXs
Peter K. Espen 6525 Applied Physics Inc.
Wed Sep 2 16:05:36 1998
> The problem is that receiving data on the linux box is approx.
> 10 times slower than transmissions. I have some perl utilities
> that allow me to measure data transfer rates without disk I/O.
> Sending data from Linux to NT I get approx. 50Mbps throughput and
> NT to Linux I am getting < 5Mbps. This difference is also
> seen in ftp transfers between NT and Linux. Ftp from NT to
> Linux is about 10 times slower than ftp from Linux to NT.
> I have run tests between NT and DEC Unix boxes at work, and
> between Sun Ultras and Redhat Linux at work and I cannot duplicate
> this difference.
> I have tried all combinations of full-duplex and half-duplex
> (by using various "options=" to the 3c59x v0.99E driver) settings
> that I can think of. I have also tried altering the settings on the
> 3c905TX on the NT side, with no changes.
OK, so I did have a cable problem that was part of the problem.
Instead of receives that are 10 times slower, they are now down
to about 2.5 times slower.
Someone had pointed out a possible discussion about this in the
linux-kernel lists, but I have yet to find such a discussion.
I would normally just chalk this up to performance characteristics
in the kernel, except for the fact that I can't duplicate this
TX/RX difference on two of the other Linux boxes I have access to.
A big difference is that the other two Linux boxes use 3Com cards
and 3c59x drivers, but are only on a 10Mbps network.