We have been using CPLEX 12.8. Now we want to upgrade to later versions. Our tests with multiple MILP instances show that CPLEX 12.8 is faster (solution time) than both 20.1 and 22.1 versions. We also tried different parameter settings but did not get any solution.

One such MILP in .lp format is attached along with the related log files. Need some insights on this. We also reported the issue to IBM ILOG support and they can reproduce the issue but we have not received any resolution yet.

The first thing to note is that when IBM develops a new version of CPLEX, they test it on a wide range of problems. Typically the new release is faster overall (so presumably on the majority of problems), but there is no guarantee that it is faster on every problem.

The second thing to note is that your sample problem generates thousands (!) of warnings due to the presence of duplicate constraints (same name, same left and right side). While CPLEX winnows those out, it may be a sign of other problems. I would advise that you check the code that generates the model.

The third thing is that CPLEX by default makes some decisions randomly, and I believe the default random number seed changes with each version. So you might want to try running the old and new versions on the same problem setting the same random seed value and see if the older version is still faster.

The fourth thing is that if you set the CPLEX data check parameter to 2 before reading and running the model, it reports that some constraints have coefficients that are very close to rational values. A sample message is "Decimal part of coefficients in constraint 'mult1_x_tanque07a_42_33001691!28!2_pct_tanque07a_42_33001691!28!2' are fractions and can be scaled with 29/1." If you rescale those constraints in the code generating the model, it might possibly make a difference.

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

Original Message:
Sent: Tue May 16, 2023 06:43 AM
From: Hamidur Rahman
Subject: CPLEX 20.1 and 22.1 running slower than CPLEX 12.8 for multiple MILP instances

We have been using CPLEX 12.8. Now we want to upgrade to later versions. Our tests with multiple MILP instances show that CPLEX 12.8 is faster (solution time) than both 20.1 and 22.1 versions. We also tried different parameter settings but did not get any solution.

One such MILP in .lp format is attached along with the related log files. Need some insights on this. We also reported the issue to IBM ILOG support and they can reproduce the issue but we have not received any resolution yet.

------------------------------
Hamidur Rahman
------------------------------

Hi Professor Paul,

Thank you so very much for your comments and suggestions. Changing the number of random seed to earlier version improves the run time for the latest versions for a couple of instances only.
For most of the other instances changing the random seed is not helping. One instance I am uploading (oss702.lp).

Just curious to know why this random seed number makes so much difference in run time.

Regards
Hamidur Rahman
Blue Yonder Inc.

------------------------------
Hamidur Rahman

Original Message:
Sent: Tue May 16, 2023 04:15 PM
From: Paul Rubin
Subject: CPLEX 20.1 and 22.1 running slower than CPLEX 12.8 for multiple MILP instances

The first thing to note is that when IBM develops a new version of CPLEX, they test it on a wide range of problems. Typically the new release is faster overall (so presumably on the majority of problems), but there is no guarantee that it is faster on every problem.

The second thing to note is that your sample problem generates thousands (!) of warnings due to the presence of duplicate constraints (same name, same left and right side). While CPLEX winnows those out, it may be a sign of other problems. I would advise that you check the code that generates the model.

The third thing is that CPLEX by default makes some decisions randomly, and I believe the default random number seed changes with each version. So you might want to try running the old and new versions on the same problem setting the same random seed value and see if the older version is still faster.

The fourth thing is that if you set the CPLEX data check parameter to 2 before reading and running the model, it reports that some constraints have coefficients that are very close to rational values. A sample message is "Decimal part of coefficients in constraint 'mult1_x_tanque07a_42_33001691!28!2_pct_tanque07a_42_33001691!28!2' are fractions and can be scaled with 29/1." If you rescale those constraints in the code generating the model, it might possibly make a difference.

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

Original Message:
Sent: Tue May 16, 2023 06:43 AM
From: Hamidur Rahman
Subject: CPLEX 20.1 and 22.1 running slower than CPLEX 12.8 for multiple MILP instances

We have been using CPLEX 12.8. Now we want to upgrade to later versions. Our tests with multiple MILP instances show that CPLEX 12.8 is faster (solution time) than both 20.1 and 22.1 versions. We also tried different parameter settings but did not get any solution.

One such MILP in .lp format is attached along with the related log files. Need some insights on this. We also reported the issue to IBM ILOG support and they can reproduce the issue but we have not received any resolution yet.

------------------------------
Hamidur Rahman
------------------------------

The first thing to note is that when IBM develops a new version of CPLEX, they test it on a wide range of problems. Typically the new release is faster overall (so presumably on the majority of problems), but there is no guarantee that it is faster on every problem.

The second thing to note is that your sample problem generates thousands (!) of warnings due to the presence of duplicate constraints (same name, same left and right side). While CPLEX winnows those out, it may be a sign of other problems. I would advise that you check the code that generates the model.

The third thing is that CPLEX by default makes some decisions randomly, and I believe the default random number seed changes with each version. So you might want to try running the old and new versions on the same problem setting the same random seed value and see if the older version is still faster.

The fourth thing is that if you set the CPLEX data check parameter to 2 before reading and running the model, it reports that some constraints have coefficients that are very close to rational values. A sample message is "Decimal part of coefficients in constraint 'mult1_x_tanque07a_42_33001691!28!2_pct_tanque07a_42_33001691!28!2' are fractions and can be scaled with 29/1." If you rescale those constraints in the code generating the model, it might possibly make a difference.

We have been using CPLEX 12.8. Now we want to upgrade to later versions. Our tests with multiple MILP instances show that CPLEX 12.8 is faster (solution time) than both 20.1 and 22.1 versions. We also tried different parameter settings but did not get any solution.

One such MILP in .lp format is attached along with the related log files. Need some insights on this. We also reported the issue to IBM ILOG support and they can reproduce the issue but we have not received any resolution yet.

Uploaded the model (oss702.lp) and the log file. It can be seen that CPLEX 12.8 has random seed = 201709013 with solution time 8.67 seconds. On the other hand, with this random seed value for CPLEX 22.1, the run time is 19.24. We are getting similar results for other MILP instances also. Please suggest what we can do.

Thanks

The first thing to note is that when IBM develops a new version of CPLEX, they test it on a wide range of problems. Typically the new release is faster overall (so presumably on the majority of problems), but there is no guarantee that it is faster on every problem.

The second thing to note is that your sample problem generates thousands (!) of warnings due to the presence of duplicate constraints (same name, same left and right side). While CPLEX winnows those out, it may be a sign of other problems. I would advise that you check the code that generates the model.

The third thing is that CPLEX by default makes some decisions randomly, and I believe the default random number seed changes with each version. So you might want to try running the old and new versions on the same problem setting the same random seed value and see if the older version is still faster.

The fourth thing is that if you set the CPLEX data check parameter to 2 before reading and running the model, it reports that some constraints have coefficients that are very close to rational values. A sample message is "Decimal part of coefficients in constraint 'mult1_x_tanque07a_42_33001691!28!2_pct_tanque07a_42_33001691!28!2' are fractions and can be scaled with 29/1." If you rescale those constraints in the code generating the model, it might possibly make a difference.

We have been using CPLEX 12.8. Now we want to upgrade to later versions. Our tests with multiple MILP instances show that CPLEX 12.8 is faster (solution time) than both 20.1 and 22.1 versions. We also tried different parameter settings but did not get any solution.

One such MILP in .lp format is attached along with the related log files. Need some insights on this. We also reported the issue to IBM ILOG support and they can reproduce the issue but we have not received any resolution yet.

Uploaded the model (oss702.lp) and the log file. It can be seen that CPLEX 12.8 has random seed = 201709013 with solution time 8.67 seconds. On the other hand, with this random seed value for CPLEX 22.1, the run time is 19.24. We are getting similar results for other MILP instances also. Please suggest what we can do.

Thanks

The first thing to note is that when IBM develops a new version of CPLEX, they test it on a wide range of problems. Typically the new release is faster overall (so presumably on the majority of problems), but there is no guarantee that it is faster on every problem.

The second thing to note is that your sample problem generates thousands (!) of warnings due to the presence of duplicate constraints (same name, same left and right side). While CPLEX winnows those out, it may be a sign of other problems. I would advise that you check the code that generates the model.

The third thing is that CPLEX by default makes some decisions randomly, and I believe the default random number seed changes with each version. So you might want to try running the old and new versions on the same problem setting the same random seed value and see if the older version is still faster.

The fourth thing is that if you set the CPLEX data check parameter to 2 before reading and running the model, it reports that some constraints have coefficients that are very close to rational values. A sample message is "Decimal part of coefficients in constraint 'mult1_x_tanque07a_42_33001691!28!2_pct_tanque07a_42_33001691!28!2' are fractions and can be scaled with 29/1." If you rescale those constraints in the code generating the model, it might possibly make a difference.

We have been using CPLEX 12.8. Now we want to upgrade to later versions. Our tests with multiple MILP instances show that CPLEX 12.8 is faster (solution time) than both 20.1 and 22.1 versions. We also tried different parameter settings but did not get any solution.

One such MILP in .lp format is attached along with the related log files. Need some insights on this. We also reported the issue to IBM ILOG support and they can reproduce the issue but we have not received any resolution yet.