next up previous
Next: Concluding Remarks Up: NetStream: a Flexible and Previous: Example of Groups Management


Performance Evaluation

In this section we present some basic performance measurements to show that NetStream is usually as fast as a raw MPI program, with a slight overhead when sending packed data. The tests have been performed in a 100 Mbps Fast-Ethernet cluster. We analyze the exchanges between two stations in this cluster (Pentium III, 700 Mhz, 128 Mb RAM).

In Figure 1 we show the time in milliseconds of sending different amount of data of the basic char", \verb"int" and \verb"double values. It can be observed that there is no overhead of NetStream over MPI for any amount of data.

Figure 1: Sending int/double values between two workstations linked by Fast-Ethernet: MPI versus NetStream times.
\begin{figure}\centering\epsfxsize =5.5cm \epsfbox{ints.eps}$\,\,\,\,\,$ \epsfxsize =5.5cm \epsfbox{doubles.eps} \end{figure}

In Figure 2 we show the trends for packets of different lengths of the basic data types int" and \verb"double as they are very usual in numerical applications. A small overhead is then detected, especially for very long packets. The user must decide whether these small delays is worth-wile in the application at hands. In a WAN environment, and e.g. for optimization tasks, the packets are usually in the region of less than 2 Kb, and the overhead can be ignored.

Figure 2: Sending packed int/double values between two workstations linked by Fast-Ethernet: MPI versus NetStream times.


next up previous
Next: Concluding Remarks Up: NetStream: a Flexible and Previous: Example of Groups Management
Enrique Alba 2001-11-15