next up previous
Next: Using the DSNIC board Up: Performance improvements Previous: Performance improvements

Modelling the CPU load

We model the CPU time to communicate a message of length n by the sum of a fixed time and a time dependent on the message length n:

The factor is the per-byte CPU time. Notice that . Knowing the CPU time to communicate a message, , and the communication latency, , the following equation expresses the CPU load :

The value , which equals , expresses the CPU load for sending a 0-byte message. Furthermore, we can derive the following for the asymptotic CPU load :

In Section 8.3, we have already used this equation. The asymptotic CPU load, , calculated there, using for the per-byte CPU time for 4096 byte packets, matches the measured CPU load for long messages.

The results presented in this paper allow us to derive the parameters of the model that match the Comms2 communication using a packet size of 1024, i.e., the typical packet size required for ATLAS, and threads to perform the computation results. From Figure 7, we can determine that , the asymptotic bandwidth, equals 16.6 Mbytes/s. Furthermore, this figure allows us to determine that . Since , we know that . Knowing from Table 3 that , allows us to determine that , using . We cannot derive from the measurements, therefore we assume the CPU load is 100 % during 0-byte message communication, so . These parameters have been used to reconstruct the performance graphs of the current system, see graph A in Figure 12.



Marcel Boosten
Wed Mar 11 14:25:07 MET 1998