Getting Started
Download CS365_Lab05.zip to your cluster head node account. Unzip it.
You will modify the source files experiment1.c and experiment2.c. You can run the command make to build the two programs.
Your Task
Your task is to perform experiments to measure two performance characteristics of our MPI cluster: latency and bandwidth.
Latency is the amount of time it takes for the first bit of a message sent from one process to be received by a second process.
Bandwidth is the rate at which data can be transferred from one process to another. Bandwidth is measured as amount of data per unit of time; e.g., bytes per second.
Use the two provided C source files, experiment1.c and experiment2.c, to perform the two experiments. You can use the runpar script to launch the programs: for example
$ ./runpar experiment1 2
will run the experiment1 program with two processes.
Hints
To measure latency, write a “ping-pong” program where process 0 sends a small (one byte) message to process 1, and then process 1 sends a small message to process 0. Process 0 can measure the amount of time elapsed from the beginning of its send to the end of its receive, and divide by 2 to find the latency.
To measure bandwidth, process 0 should send a large volume of data (at least 1 megabyte) to process 1. Process 1 should send a small acknowledgement message to process 0 when all of the data has been received. Process 0 can measure the bandwith by measuring the amount of time that has elapsed from when it starts sending the data to when it receives process 1’s acknowledgment. Note: each MPI_Send operation should send a relatively large amount of data (e.g., 128 KB) to reduce the effect of per-message communication overhead.
Timing
You can use the utime function to measure time with up to microsecond (millionth of a second) accuracy. Use it as follows:
unsigned long start, end, elapsed;
start = utime();
code you want to benchmark
end = utime();
// Find total number of microseconds elapsed
elapsed = end - start;