High Performance TCP/IP communications with HPs and IBMs connected around an Ancor Fibre Channel Fabric
The program used to test out the performance is `netperf'. This program is in the public domain and has been originally evolved by Rick Jones (raj@cup.hp.com, HP Lab in Cupertino, California, USA). This program can be obtained by anonymous ftp to iworks.ecn.uiowa.edu. In [1] and [2], communication benchmarks were achieved using `ttcp' program. The reason why we used `netperf' is because of a more flexible and more convenient user interface and because of a stronger potentiality. Otherwise, there is no difference in the performance issued from the one or the other.
We used five input parameters to this program, which can be chosen independently:
The duration length for the test was equal to 30 seconds, in order to provide relevant measurements.
Several values for the socket buffer size have been tested according to the experiments and to the machine specifications: 131072 bytes, 262144 bytes, 349525 bytes, 524288 bytes and 1048576 bytes.
When the TCP_no_delay option is not set, messages are grouped before being sent. It does not affect particularly the bandwidth estimation, which is generally measured for large messages. It is completely different in the case of evaluating the overhead of the system (interface plus fabric plus all the software): the measurement will indeed return an over optimistic expectation. The influence of grouping messages is of course more sensitive on small messages than on large messages. To allow a realistic overhead measurement, the TCP_no_delay option must be turned on. In that case, it will ensure the sender to send packets one by one.
Graphs show hereinafter how the communication time and therefore the bandwidth scale with the size of the message to be transferred and the socket buffer size. Figures a display bandwidth measurements, whilst figures b display corresponding communication times.
Generated with WebMaker