Homework 1
Assigned 15 Feb 2006
Due 22 Feb 2006
20 points


Problems 1-1, 1-2, and 1-3, page 41.

Problems 2-1, 2-3, page 77


Problem M1:

Suppose processors are numbered 1 to N, for N a power of 4, and
each processor simultaneously sends a message to a destination
that is some permutation of 1 to N.  Compare the effectiveness
of a hypercube, 2D grid, and Omega network for this situation.


Problem M2:

Write a short MPI program to measure the approximate message
latency between two blades.  One way to do that is to bounce 
a short message back and forth repeatedly and time the overall
result with the unix "time()" call.  time() is only accurate
to a second, so you will need a few thousand bounces to get 
a meaningful measure.  If a finer resolution timing call is
available for the blade C libs you can use that and fewer
bouncs, if you'd prefer.  In either case, print your result 
in microseconds for a typical message transfer.

Compare your latency results with what you get from ping.  Run
"flops" on blade33, and see what you get.  Compare the cost of 
a local and remote floating point operation.