Decision Optimization

Decision Optimization

Delivers prescriptive analytics capabilities and decision intelligence to improve decision-making.


#Analytics
#DecisionOptimization
#DecisionOptimization
 View Only

  • 1.  CPLEX 22.1 performance is slower than 20.1

    Posted 07/26/22 10:17 AM
    Hi, all. In the latest release of our software we've upgraded CPLEX from 20.1 to 22.1 and noticed significant time increase when executing identical simulation with identical CPLEX settings between our previous version (with CPLEX 20.1) and latest one. I believe CPLEX settings are mostly default except for the Relative Gap. It's set to 0.8 for this particular client. Any suggestions which settings are worth tweaking to speed up 22.1 performance?
    I appreciate an advice from the experts.

    ------------------------------
    Marina Steinberg
    ------------------------------

    #DecisionOptimization


  • 2.  RE: CPLEX 22.1 performance is slower than 20.1

    Posted 07/26/22 11:35 AM
    In a side by side comparison of node logs (using 20.1 v. 22.1), can you identify a specific change? Is the node throughput (nodes per second) worse with 22.1? Is 22.1 slower improving the best solution, or slower improving the best bound?

    ------------------------------
    Paul Rubin
    Professor Emeritus
    Michigan State University
    ------------------------------



  • 3.  RE: CPLEX 22.1 performance is slower than 20.1

    Posted 07/27/22 05:59 AM

    Thank you, Paul, for your response. Below you can see examples of times recorded by our log which are slower in 22.1. Looks like it is slower improving the BIS and i don't think that the node throughput is reported for us. I am also including a screenshot of our settings - please take a look - maybe there's a setting that we ignore but could optimize performance? Thank you for all the help!

    Root node processing (before b&c):
    Real time = 28.72 sec. (13791.13 ticks)
    Total (root+branch&cut) = 28.72 sec. (13791.13 ticks)
    Presolve time = 1.34 sec. (570.56 ticks)
    Found incumbent of value 3.1115092e+11 - This is BIS value! - after 1.75 sec. (796.01 ticks)
    Probing time = 0.11 sec. (31.71 ticks)



    ------------------------------
    Marina Steinberg
    ------------------------------

    Original Message


  • 4.  RE: CPLEX 22.1 performance is slower than 20.1

    Posted 07/27/22 11:31 AM
    Marina,

    Sorry, but just seeing the presolve and root probing stats from one of the two solvers does not tell us anything. What we would need to see would be the full logs (or at least a significant chunk), starting with the root node processing and including a decent amount of the branching ... and we would need to see the logs from both 20.1 and 22.1, so that we could look for differences.

    As for node throughput, that is not reported directly but can be inferred from the logs. The logs indicate cumulative node count and periodically contain a line indicating elapsed time. Putting those together, we can get a ballpark estimate of nodes crunched per second.

    Paul

    ------------------------------
    Paul Rubin
    Professor Emeritus
    Michigan State University
    ------------------------------

    Original Message


  • 5.  RE: CPLEX 22.1 performance is slower than 20.1

    Posted 07/27/22 05:42 AM
    Hi,

    It's possible that 22.1 is slower than 20.1 on this particular instance.

    Have you tried https://www.ibm.com/support/pages/cplex-performance-tuning-mixed-integer-programs ?

    Have you tried to set the MIP Emphasis switch https://www.ibm.com/docs/en/icos/22.1.0?topic=parameters-mip-emphasis-switch to 5 ?

    regards


    ------------------------------
    [Alex] [Fleischer]
    [Data and AI Technical Sales]
    [IBM]
    ------------------------------



  • 6.  RE: CPLEX 22.1 performance is slower than 20.1

    Posted 07/29/22 05:38 AM
    Thank you, Alex!!! Running with MIP Emphasis switch set to 5 (heuristic) indeed decreased the runtime making it even slightly better than in previous version. Greatly appreciate your suggestion!

    ------------------------------
    Marina Steinberg
    ------------------------------

    Original Message


Related Content