A severe memory leak exists when executing an algorithm multiple times

[svonolfe]

When executing an algorithm multiple times, the memory consumption skyrockets. To illustrate this phenomena, try the attached test project which executes the GA_VRP algorithm multiple times (using 10 iterations).

I was able to trace back the problem to the Run::Initialize method using the CLR profiler and created a patch for that method. The test project contains the patched dll.

Please note if you enable algorithm.StoreAlgorithmInEachRun you can see the memory leak. After applying the patch, if you disable algorithm.StoreAlgorithmInEachRun the memory leak seems to disappear, or at least becomes less severe.

The problem seems to be related to the recent cloning refactoring.

[swagner]

Worked on in r4825.

[mkommend]

Tested the implemented changes in the CLR profiler and the memory consumption dropped dramatically.

[mkommend]

The memory leak regarding the parameters is fixed as far is I can see in the CLR profiler. Although a memory leak still exists while collecting the result values, but I could not exactly track down what causes this memory leak.

I still have to code review the changes made in r4825 and every problem / algorithm that does no full deep clone must be corrected.

[abeham]

To the best of my knowledge, algorithms, at least the ones I implemented, perform a full deep clone.

I remember from the cloning refactoring, that the VRP problem does not fully deep clone.

[mkommend]

Corrected cloning of results and parameters in Run with r4833.

[mkommend]

Replying to mkommend:

Although a memory leak still exists while collecting the result values, but I could not exactly track down what causes this memory leak.

This memory leak has the same cause as the first one. Results get collected and the ResultsCollection registers on ToStringChanged and ItemImageChanged. If a result is the same as a parameter of the problem or the algorithm (e.g. BestKnownQuality) the whole ResultCollection can't be released and collected by the GC. A resolution to this problem would be if all collected values get cloned before they are collected (ticket #1279).

[mkommend]

Corrected memory leak in Problems.DataAnalysis with r4834.

[mkommend]

Reviewed and tested the changes from r4825.

[swagner]

According to the changes described in this ticket, please adapt the VRP, if it contains any parameters which are not deeply cloned.