Originally posted by: ChristopheV

Hi all,

I recently experienced some out-of-memory problems while solving a big industrial MIP problem (problems mainly arise during polishing), and for that reason I've tried to migrate from 32-bit to 64-bit. To my surprise, I've noticed that the 64-bit version is working slower, however, than its 32-bit equivalent.

To illustrate the slow-down: to solve exactly the same problem, it takes 7.4 sec to solve the root node in 32-bit, while it takes 12.98 sec on the 64-bit machine. In the same amount of time, 1903 MIP nodes were visited in 32-bit, while the 64-bit version only visited 37 nodes.

Of course, the test were performed on a different computer, but, based on their specifications, I expected the 64-bit machine to be faster. On the 32-bit machine, for instance, I have a 2.33GHz Intel Xeon CPU, while on the 64-bit machine, I have a 3.00GHz Intel Xeon CPU. The 32-bit machine contains 3.25GB of RAM (and I am using the 3GB-switch), while the 64-bit machine has 5GB of RAM. I realise that there are other parameters that may influence the speed of a computer, but in all aspects, the 64-bit machine seems to be "better".

I basically have two questions:

1) Did others experience a slow-down of CPLEX while migrating from 32-bit to 64-bit?
2) If so, is it normal that the 64-bit version works slower? This is something I wouldn't expect intuitively.

Thanks a lot!
Christophe
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: SystemAdmin

I agree, this sounds kind of unexpected. Could you provide log files for the two different runs? Maybe they can give a clue about the differences.
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: ChristopheV

Hello,

thank you for your reply.

You will find the log files in attachment. I used normal level of logging (i.e., not the most detailed level) in this case. As you will see, I used 8 threads on the 32-bit machine and 4 on the 64-bit machine, but this does not influence the results I get. So if I use 4 threads on the 32-bit machine, it still takes about 8 seconds for the root node to be solved.

Thanks again,
Christophe
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: SystemAdmin

There already is a big difference in the "Root relaxation solution time" (7.44 vs 12.98 sec). Since you are using opportunistic parallel (explicitly setting the "threads" parameter automatically switches to opportunistic parallel), you are getting a concurrent LP solve for the root LP. My guess is that this, combined with the different threads settings and potentially with different memory bandwidths in the machine can explain the difference in the root LP solve time. The fact that the the initial root LP objective values and iteration counts do not match supports the assumption that the concurrent LP root solve does different things on the two platforms.

Could you please try the following experiments:
1. Set the "MIP display parameter" to 4, such that we can see the root LP iterations. Also set "threads" to 4 in both runs, and set the "parallel mode" to 1 ("set parallel 1" in the interactive shell). Then rerun your test.
2. The same as in (1), but set "threads" to 1.
Thanks,

Tobias
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: ChristopheV

Dear Tobias,

thank you for your answer. I think your suggestions are very interesting, and I will try them as soon as I have access again to the 64-bit server (it is a machine that is shared amongst many people, and last time, we only had a limited time window for testing purposes). I'll post my results as soon as I have them.

Thanks!
Christophe
#CPLEXOptimizers
#DecisionOptimization