pingpongSweep

Does a ping pong test between all pairs of allocated nodes. In general, this should be run with one task/rank per host so that ping pongs aren't sent between processes on the same physical host.

Note: these are logical nodes from the Charm++ perspective, which are really processes. When running in SMP mode, one has +p / ++ppn number of nodes. In non-SMP mode, one has +p nodes (as each PE is an independent process)).

Usage

./pingpong <upper limit of message size in bytes> <number of iterations> <number of warmup iterations>

For each pair of nodes, sweep through message payload sizes from 0 bytes to the specified upper limit (1,024 bytes by default), doubling each time, and for each size, do the specified number of round trip ping pong iterations (10,000 by default plus 100 warmup iterations), then report the average time each round trip took.

Or, in pseudocode:

for i in 0..numNodes
  for j in i + 1..numNodes
    for size in 0..maxSize // 0, 1, 2, 4, ..., maxSize
      startTimer
      for iteration in 0..numIterations
        send message of specified size from node_i to node_j and then back to node_i from node_j
      endTimer
      report average roundtrip time, or (endTimer - startTimer) / numIterations

Output

The program outputs the average time in microseconds, parenthetically followed by the message size and the indices of the two nodes involved).

For example, here is the output corresponding to a message payload of size 0 for the ping pong between node 0 and node 1:

Avg roundtrip time is 5.423577 us (size 0 (0, 1))