On Sat, May 13, 2000 at 11:53:31PM -0700, Ben Greear wrote:
> I'm having some interesting problems with packet loss while
> trying to run a user-space protocol directly over ethernet. Basically,
> when trying to do about 64kbps (full duplex) on 10 socket connections, I'm
> losing about .1% of the packets. The intervening network is a cheap
> 10/100bt switch. The ethernet cards are two 4-port ZNYXs (tulip) on one, and
> LinkSys EtherFast card on the other computer.
>
> If I move the rate down to 24kbps, then the packet loss seems to dissappear.
>
> I'm wondering if the problem is because the packet buffers in the
> raw socket are too small and the user-space program is not able
> to read them before some pkts get flushed?
You can increase the kernel buffers with the SO_RCVBUF/SO_SNDBUF setsockops
and the net/core/[rw]mem_{max,default} sysctls. For large values you
may also need to increase the vm/freepages settings (see the man pages
for more details)
There are unfortunately not counters for all oom packet drop cases,
so some get silently dropped. It may also be dropped somewhere else
(like the switch).
-Andi
-
To unsubscribe from this list: send the line "unsubscribe linux-net" in
the body of a message to [EMAIL PROTECTED]
