Parallel.For Slows Down with more cores

Question

Hi,

I'm running my test app on Intel Xeon 514 / 2.33 GHz, 2 Sockets, 2 Cores each, 4 GB RAM

the test app is a Monte Carlo Scenario that creates 500.000 objects of Type Mandelbrot; each object creates random values and forecasts 20 days; the result is 500.000 x 20 random values

the code looks like:

for (int CurrentDayOfForecast = 0; CurrentDayOfForecast < ForecastRange; CurrentDayOfForecast++)
{
    #region Calculate Current Day
    Parallel.ForEach(Mandelbrot, delegate(_30DaysUpDown_II_Mandelbrot CurrentMandelbrot)
    {
        CurrentMandelbrot.Forecast(CurrentDayOfForecast);
    });
    #endregion
}

when I run this code on 1 Socket (2 cores) the server is utilized at 45-48% and the code finishes after 40 seconds
when I run this code on 2 Sockes (4 cores) the server is utilized at 96-98% and the code finishes after 55 seconds

so more cores = slow down


Any hints?

Answer 1

Are the all the tasks reading and updating a shared piece of memory.  This can often result in cache lines constantly being invalidated and such cause sub optimal performance as the cores are having to keep rebuilding their caches.

Answer 2

In addition not necessarily shared data, but data close together in memory such that they occupy the same cache line.

Answer 3

In addition to Andy's good questions, have you tried enabling the server GC?  If not, you should try that as well, by creating a .config file for your application containing:

<?xml version="1.0" encoding="utf-8" ?>
<configuration>
  <runtime>
    <gcServer enabled="true"/>
  </runtime>
</configuration>

Answer 4

Cool!

It speeds up calculation. Not that much but it's faster...

Answer 5

If you can, you need to invert the outer and the inner loop.
The outer loop is the one that needs to be parallelized.
The inner loop is maybe fast enough. So much fast than injecting Parallelism causes the execution to be slower.

Parallel.For(0, ForecastRange, CurrentDayOfForecast =>{
  foreach(var mnd in Mandelbrot)
  {
     // Do something with mnd
  }
});