Parallella Community

tery, tony, xilman:

Thanks for the feedback! Three variables have changed: new defconfig, updated linux kernel, and using a newer linaro tree. I hope to have a new beta headless release version up this week for some brave souls to try.

Right now, it does look like the new configuration has possibly boosted ethernet performance from ~200Mb/s to 500Mb/s. I haven't found the root cause for this nice bump, so fingers cross that it's actually "true".

Andreas

Alternatively use the venerable utility and you can compare your throughput against the nice little table of that we created at Element14 for the most common current ARM boards, plus a few others for comparison.

Details of running nuttcp are given in the linked thread, but in a nutshell it's simply to run "nuttcp -S" on some server "X" and then run "nuttcp -t X" and "nuttcp -r X" as transmit and receive tests to and from the server. Parallella can play the server or client role, whichever is more convenient. (iperf should of course give similar results, although comparing against our table is most fairly done using the same utility we used.)

The maximum theoretical TCP throughput over gigabit Ethernet is 941.482 Mbps with TCP TimeStamps enabled, or 949.285 Mbps without. All modern x86/amd64 equipment reaches that limiting throughput without any difficulty at all, but the gigabit-capable ARM boards we tested were nowhere near the theoretical maximum by quite some distance. Some boards couldn't even fully utilize a 100 Mbps link.

Hopefully this year will see newer ARM devices putting up a better showing.

Morgaine.