em blues

[Danny Braniss]

> On 10/11/06, Danny Braniss <danny at cs.huji.ac.il> wrote:
> > the box is a bit old (Intel Pentium III (933.07-MHz 686-class CPU)
> > dual cpu.
> >
> > running iperf -c (receiving):
> >
> > freebsd-4.10    0.0-10.0 sec    936 MBytes    785 Mbits/sec
> > freebsd-5.4     0.0-10.0 sec    413 MBytes    346 Mbits/sec
> > freebsd.6.1     0.0-10.0 sec    366 MBytes    307 Mbits/sec
> > freebsd-6.2     0.0-10.0 sec    344 MBytes    289 Mbits/sec
> >
> > btw, iperf -s (xmitting) is slightly better
> > freebsd-4.10    0.0-10.0 sec    664 MBytes    558 Mbits/sec
> > freebsd-5.4     0.0-10.0 sec    390 MBytes    327 Mbits/sec
> > freebsd-6.1     0.0-10.0 sec    495 MBytes    415 Mbits/sec
> > freebsd-6.2     0.0-10.0 sec    487 MBytes    408 Mbits/sec
> >
> > so, it seems that as the release number increases, the em
> > throughput gets worse - or iperf is.
> 
> You arent measuring em, you're measuring RELEASES on
> your hardware, is this a surprise on a P3, no.
> 
I agree 100% with your first statement, but, if_em is useless without
the rest, and if it's not delivering, then something is wrong somewhere,
no necesarely with the em driver, but in how the system interacts.

> I still do 930ish Mb/s on a P4 with a PCI-E or PCI-X adaptors
> running 6.1, in fact can do that with a 4 port adaptor I believe.

i do get on certain combinations nice numbers:
Server listening on TCP port 5001
TCP window size: 64.0 KByte (default)
------------------------------------------------------------

[  4]  0.0-10.0 sec  1.08 GBytes    928 Mbits/sec
(the mb is Intel SWV).

which seems almost optimal, but on other platforms i get
[  3]  0.0-10.0 sec    654 MBytes    548 Mbits/sec
(the mb is Intel SE7501)

cheers,
	danny