FEC performance on the 855T/860T

FEC performance on the 855T/860T

[Graham Stoney]

Hi gang,

Thanks to all your help, our custom 855T hardware is now up and running, and
it's time to do some system profiling & tuning. The first bottleneck to stand
out is that the FEC port isn't going as fast as we'd like. A simple socket
based tcp client/server setup which repeatedly send's from another machine and
simply recv's data without processing it in user space at all only gets just
over 3 Mbytes/s between two machines connected via a 100 Mbps switch.

There was some discussion on the mailing list last December regarding FEC
throughput on the 860T. For example, see:
    http://lists.linuxppc.org/listarcs/linuxppc-embedded/199912/msg00061.html

Our Embedded Planet CLLF860T boards fair even worse, getting around
2.7 Mbytes/s over the FEC, which tells me that it's not a problem with our
hardware. This is less than a third of its theoretical peak, and I'm
wondering if anyone has had a closer look at where the bottleneck with the
FEC lies?

Regards,
Graham
--
Graham Stoney
Principal Hardware/Software Engineer
Canon Information Systems Research Australia
Ph: +61 2 9805 2909  Fax: +61 2 9805 2929

** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

Re: FEC performance on the 855T/860T

[Dan Malek]

Graham Stoney wrote:


> .....The first bottleneck to stand
> out is that the FEC port isn't going as fast as we'd like.


We have discussed this before.  The FEC runs at 100 Mbits/sec,
back to back packets.  It has to....there isn't any buffer
space in the FEC to do otherwise.  The bottleneck is the protocol
and application processing in the PPC core.  The performance of
a TCP/IP connection scales nearly linearly with the processor
speed.

The 50 MHz 860P will run about 30 Mbits/sec TCP/IP, and a 100
MHz 860P will run about 60 Mbits/sec TCP/IP.  You can write
a program that will send back-to-back Ethernet frames at the
maximum rate, and will read and discard Ethernet frames at the
maximum rate.

Just write a program (nttcp works) to benchmark using the local
loopback device.  That will tell you what the processor can
accomplish.


	-- Dan

** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

TCP/IP performance on the 855T/860T

[Graham Stoney]

Dan Malek writes:
> The bottleneck is the protocol and application processing in the PPC core.

So has anyone had a chance to look at where exactly in the TCP/IP stack the
bottleneck is since this was last discussed?  Surely a 50 MHz part with an
on-chip 100 Mbps Ethernet controller has enough grunt to keep up with a single
TCP/IP connection.

> Just write a program (nttcp works) to benchmark using the local
> loopback device.  That will tell you what the processor can
> accomplish.

Thanks for the suggestion. I ran my tcp test through the loop device on the
50 MHz 860T CLLF board, and it gives a little over 4 Mbytes/sec. So Dan's
absolutely right as usual; the problem isn't in the FEC, it's in the protocol
stack. Somewhere. Any suggestions?

Thanks,
Graham
--
Graham Stoney
Principal Hardware/Software Engineer
Canon Information Systems Research Australia
Ph: +61 2 9805 2909  Fax: +61 2 9805 2929

** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

Re: TCP/IP performance on the 855T/860T

[Dan Malek]

Graham Stoney wrote:

> So has anyone had a chance to look at where exactly in the TCP/IP stack the
> bottleneck is since this was last discussed?

Yes, you can do a few things with TCP/IP windows to help a little,
but it is only a few percent.

> ....  Surely a 50 MHz part with an
> on-chip 100 Mbps Ethernet controller has enough grunt to keep up with a single
> TCP/IP connection.

There isn't any other processor that will do it at that speed.
Dig out an old 486 and try the local loopback test......You wouldn't
be surprised if it can't keep up.

>  Any suggestions?

...8260....



	-- Dan

** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/