Updating ExecutionTime in Engine is not thread-safe

[swagner]

The execution time of an Engine is updated using a System.Timers.Timer which processes its Elapsed events using worker threads from the thread pool. If the system is under high load and many worker threads are currently active, it might happen due to context switches that processing a single Elapsed event takes longer than the timer interval. In this case the Elapsed events overlap which results in wrong execution time values.

[swagner]

In ​http://msdn.microsoft.com/en-us/magazine/cc164015.aspx the differences of .NET's timers are discussed. The problem of timer event reentrance is also covered at the end of the article.

[swagner]

Implemented solution as described in the article and incremented timer interval to 250ms in r5240.

[cneumuel]

The ExecutionTime is now correctly updated even in my Meta-Optimization experiments.

My performance tests show that increasing the update-interval to 250ms was already a pretty good improvement.

The following ExecutionTimes are from small Meta-Optimization Runs with ParallelEngine. Each evaluation in Meta-Optimization also executes an algorithm which updates its own ExecutionTime. The runs were executed on a blade node with 8 cores, so approximately 9 threads were updating their ExecutionTime concurrently.

	100ms	250ms	1000ms
Run 1	01:06:35	00:36:41	00:30:36
Run 2	00:48:16	00:31:42	00:34:22

[swagner]

Prevented ObjectDisposedException which was thrown in Engine when accessing an already disposed timer in r5444.