Next: Performance Model
Up: Load Balancing in
Previous: Load Balancing in
The hypothetical best machine is the one yielding the smallest
execution time T for a given problem P. Therefore, we have to
estimate this time on every machine M in the NetSolve
system. Basically, we split the time T into and , where
- is the time to send the data to M and receive the result over the network, and
- is the time to perform the computation on M.
The time can be computed by knowing
the following
- network latency and bandwidth between the local host and M,
- size of the data to send, and
- size of the result to be received.
The computation of involves knowledge of
the
- size of the problem,
- complexity of the algorithm to be used, and,
- performance of M, which depends on
- the workload of M and
- the raw performance of M.
Joint Institute for Computational Science
Mon Apr 29 13:00:40 EDT 1996