-
Notifications
You must be signed in to change notification settings - Fork 129
NIC's Benchmark
netperf
To set up this benchmark please visit www.netperf.org.
Specific instructions for the version we are using (2.5.x) are here in the section: Installing-Netperf-Bits
Scripts presume netperf is installed in /usr/local/bin
- tcp_stream_script
Download: tcp_stream_script.sh
- udp_stream_script
Download: udp_stream_script.sh
On Machine A
/usr/local/bin/netserver
This will spawn a netserver server sufficient for all remaining tests.
On Machine B
/usr/local/bin/netserver
and then presuming you're in
./[tcp|udp]_stream_script.sh [machine A's IP] [CPU] [-Tx,x] > [udp|tcp]_[testname].txt
This will run the udp_stream_script and tcp_stream_script tests between machine A and B. Here [CPU] flag displays the host/remote CPU utilization and [-Tx,x] is used to assign host/remote processor affinity.
Then simply upload the udp_.txt and tcp_.txt files to a summary of your performance results here.
The throughput of the 10G reference NIC has been compared to a commercial 10G NIC from Myricom. We measured throughput of about 1.5Gb/s using the netperf UDP stream test, with NetFPGA-10G connected directly to our Myricom 10G NIC. (For comparison, in a back-to-back test, a pair of identical Myricom 10G NICs were able to achieve at least 4.5Gb/s.)
A simple round-trip UDP latency test carried out using UDP packets measured an average latency of approximately **180us **for the NetFPGA 10G Reference NIC, with two 10G interfaces looped back using a cable. Using the same methodology, an average round trip latency of around 30us was measured for the Myricom 10G NIC.
More detail on these measurements can be found in the netperf test results for Xilinx (Dublin).