Using Vampir to Show Load Imbalance

First we select a display of MPI events only. No surprise that MPI_Bcast is the dominant activity in two of the three nodes; there is load imbalance going on here. [Pie Charts show load imbalance]

This is seen more clearly in the corresponding bar chart: [Bar Charts show load imbalance]

Node 0 is doing most of the MPI_Recv while nodes 1 and 2 are doing most of the MPI_Bcast.

Possible causes:

  1. We are printing distributed data (node 0 prints by default)
  2. There is load-imbalance (VT shows this to be the case)
(Back to main talk)