> It seems that you're not aware that people are already doing ATM and
> getting reasonable throughputs (apparently mainly limited by memory
> bandwidth).
>
> We may have too small receive queues (I think they default to something
> like ten packets), but I doubt we have any fundamental problems.
I said, it is not widely known fact.
On 100-150Mbit interface this boundary is hit by ~128byte packets.
That's why it is known only to people who make benchmarks.
This problem show on surface only on routers and with Gbit rates.
We process one packet ~10usec, NOT taking into account
copying to user and protocol processing. It means, that we drop
100% the packets, when receive more than 100Kpps.
And length of receive queues (BTW it is 300 packets) does not
affect this behaviour at all.
The simplest test, which anyone can make, is spraying 100Mbit
interface with <128byte packets on 200MHz Pent.
We spend 100% of time in driver's *_interrupt routine, and
drop all the packets in this case and making no more work.
Even timer does not tick 8)
Alexey Kuznetsov
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/