Hi!
On 14:31 Tue 21 Apr , Jeffrey Cao wrote:
> On Tue, Apr 21, 2009 at 06:49:13AM +0200, Michael Blizek wrote:
> >
> > > CPU causes problems only when you test with small packets, say 64 bytes
> > > packets. Small packets will cause more frequent interrupt for the CPU.
> > > If you can pass 24 Mbps with 2k data size, then you at least can pass
> > > 24 Mbps with bigger packets.
> >
> > If you use UDP, the kernel has to fragment and concat big packets on IP
> > layer.
> > At least in theory, this can cause CPU usage to increase.
> >
>
> I don't think the package fragement is an issue to cause CPU usage to increase
> greatly. That's a simple computation work, I don't think that will cause much
> CPU time. What affect greatly is the hardware IRQ caused by each packet
> arrive
> and the soft IRQ for each packet, and the kernel/user space switch.
>
> If you did some practical network performance test work, you'd have
> understood
> that bigger packet size will cause better througput performance. Only smaller
> packets will cause trouble.
You are right - small packets are probably way more often the cause of low
performance than "oversized" packets. However, I am not talking about transfer
of large amounts of data via TCP, but about big UDP packets. They are quite
rare and all I said was "At least in theory, this can cause CPU usage to
increase".
-Michi
--
programing a layer 3+4 network protocol for mesh networks
see http://michaelblizek.twilightparadox.com
--
To unsubscribe from this list: send an email with
"unsubscribe kernelnewbies" to [email protected]
Please read the FAQ at http://kernelnewbies.org/FAQ
