CODEWITHSUNDEEP
c++network-programmingmpi
Daniel
Daniel

Reputation: 2726

How to measure the network latency in a parallel C++ program?

I am using a C++ code which uses mpi to do the parallel computing on a HPC. I want to know how to implement in my code to measure the network latency? (I want to find out if the latency is a big issue or not in my simulation.)

Does cpu's utime or stime include this latency?

PS: My cluster is using Linux, managed by SGE.

Upvotes: 3

Views: 1193

Answers (2)

JamesTullos
JamesTullos

Reputation: 242

If you just want to measure your latency, time how long it takes to send a 0 length message, or send many and average it. There are several benchmarks available as well. I work for Intel, supporting some of our Software Development Products. One of those is the Intel® MPI Benchmarks, a set of open source MPI benchmarks you can use with any MPI implementation. If you want to give it a try, you can download it at https://software.intel.com/en-us/articles/intel-mpi-benchmarks/.

Upvotes: 0

niklasfi
niklasfi

Reputation: 15931

You should have a look at VampirTrace

VampirTrace screenshot

It allows you to track the the latency as well as the thoughput of mpi calls as well as many other metrics. Since it is a commercial license you do need a license for it. I would reccomend you ask your the administrator of your cluster whether your institution has a license (for universities and the like this is often the case).

Upvotes: 2

Related Questions