Need for Timing Information

An obvious question is the need for storing timing information. Timing information mainly tells us about the communication needs, or network needs over periods of time. This information combined with things like avaiable bandwidth on the link and other communication parameters tells us of the communication patterns over the different links of the cluster. So we can determine the load on the different links over the period of program execution. It also tells us of communication wait dependencies. We can explain unseen wait periods for nodes.

In the end this allows us to identify deficiencies in the current task mapping scheme and lets us overcome them. We can thus have more efficient task mapping as a result of this.

There is also an obvious overhead associated with the whole setup of recording, storing, relaying and replaying using the timing information. Moreover the overheads are in time, space and communication. Towards the end in 7 we try to analyze this overhead in a subjective manner for the implementation.