debugging and visualization

In lecture 7 we close chapter 2 on message-passing computing, introducing the nonblocking MPI_Isend/MPI_Irecv and corresponding MPI_Test and MPI_Wait. As a motivation, we continue our exploration of the computation/communication ratio of lecture 6, asking the question: how many calculations can we perform while sending a large matrix?

The program "cost_one_sr.c" sets the time to send a 2520-by-2520 matrix of doubles to about 5 seconds. In that time frame, the manager can do one million additions of doubles, as is done in "cost_i1sr.c".

The second part of this lecture concerns MPE, a tool for a postmortem analysis based on trace files.

Bibliography