Skip to main content

Big variance of performance with same randomseed

Answered
Walter Gisler
  • Gurobi-versary
  • First Question

I thought this would be an interesting question for the community as it could also benefit others.

We were (by accident) running the same model on two separate machines (both machines have the exact same hardware specification, same version of Python, gurobipy and Gurobi) with the same random seed. It caught my attention that one of the machines didn't find a solution at all after an hour, while the other machine found an initial solution after just a few minutes.

I decided to test if I could replicate this. After a software update and restarting both machines, we were running the same model several times with the same random seed on both machines. The result was somewhat surprising: both the time to find an initial solution as well as the mip gap after an hour was varying considerably (even on the same machine), as you can see in the attached screenshot.

I had assumed that given the same randomseed, number of threads and hardware (and no other things running at the same time), we should be able to expect roughly the same performance (+/- a few percent).

Any experience with this kind of problem? Maybe a hardware problem? (the servers are about 4 years old and have been pushed to their limits on a consistent basis, so it can definitely be that there is a problem with cooling/ overheating). Unfortunately I don't have physical access to the machines right now, but I did run the Intel Processor Diagnostic Tool and everything looked normal. The CPU is a Xeon E5-2690 v4 and both servers have 128 GB of memory (which should have been more than enough in this case).

0

Comments

2 comments

Date Votes
  • Official comment
    Simranjit Kaur
    • Gurobi Staff Gurobi Staff
    This post is more than three years old. Some information may not be up to date. For current information, please check the Gurobi Documentation or Knowledge Base. If you need more help, please create a new post in the community forum. Or why not try our AI Gurobot?.
  • Matthias Miltenberger
    • Gurobi Staff Gurobi Staff

    We are going to discuss this in an internal ticket.

    Thanks,
    Matthias

    0

Post is closed for comments.

Didn't find what you were looking for?

New post