MSDN Magazine - December 2008 - (Page 19)

One possible approach to address cation is parallelized, the results may this problem is to incorporate some be computed in a different order by sort of dynamic-tuning logic to adjust the concurrent tasks. Then it is up to concurrency levels based on the apthe programmer to sort out the results plication (for example, using statistiwhere necessary to ensure correctness cal methods). Further inspection of and functionality. concurrency control and tuning apAdjusting programs to perform in proaches is also out of the scope of concurrent and parallel environments our present discussion. is not trivial. General-purpose tuning Appropriate tuning may lead to is even harder, and it cannot guarantee improvements in performance. The optimal behavior for all apps. Underfollowing is an encouraging example standing the underlying structure of showing a similar experiment, where the concurrent environment can help many factors and parameters in the you determine which parameters to Time (Seconds) tune for and the best programming ThreadPool were manually tuned speFigure 4 Improvement Due to Manually Tuning the strategies to employ. Ongoing work cifically for this test application. ThreadPool The performance results are in Figfor future ThreadPool versions inure 4. The improvement (represented by the red line) is mainly a cludes trying to reduce the need to do manual tuning. result of adding more threads, which quickly increases the concurrency levels. In the ThreadPool (represented by the blue line), only ERIKA FUENTES, PhD, is a Program Manager on the CLR team where she works one thread is added when increasing concurrency levels is deemed on the Core Operating System in concurrency. She has several academic publications in the area of statistics, adaptive systems, and scientific computing. necessary—in this experiment the check happens every interval of ERIC EILEBRECHT is a Senior Software Development Engineer in the CLR team, 250ms. Moreover, after a certain point it concludes adding more where he works on the Core Operating System. He has been at Microsoft for 11 threads because there won’t be any further improvement. years and had worked on back-end storage systems for Hotmail prior to joining In this experiment, WorkItems are queued, each of which does the CLR team. CPU work then blocks for some amount of time. Every 30 seconds the blocking time is changed: in the first 30 seconds it lasts 1ms, then 2ms, then it switches back to 1ms. At the 90 second mark, the work is stopped for 10 seconds. After this, the previous blocking cycle continues. The blue line shows performance results using the existing ThreadPool. The red line shows the improvements obtained with manual tuning of this ThreadPool. Throughput (Completions/Seconds) Build interactive diagrams easier than you ever imagined A Final Note on Extracting Parallelism The introduction of parallelism can improve performance but is not trivial to incorporate parallel concepts and algorithms into existing applications. (For example, there are many pitfalls, as outlined in “Concurrency Hazards: Solving 11 Likely Problems in Your Multithreaded Code” at msdn.microsoft.com/magazine/cc817398.) The Task Parallel Libraries (TPL) are extensions to the .NET Framework, developed by the Parallel Computing Platform team, that provide a way to expose concurrency and provide support to take advantage of parallelism. The TPL are currently available, together with more information about this development, at msdn.microsoft.com/concurrency. C reate custom interactive diagrams, network editors, workﬂows, ﬂow- charts and design tools. For web servers or local applications. It’s easy-to-use and very ﬂexible. We’re the ﬁrst developer of diagramming components and still the best. Find out for yourself: download our FREE fully functional evaluation kit, with full support at www.nwoods.com. Of course, taking advantage of exposed parallelism is still not seamless for the average programmer since there are still many factors that need to be taken into consideration. And as we saw in Figure 1, applications don’t always scale properly. The same is true when parallelizing a problem using divide-and-conquer techniques (recursive by nature). This can be tricky because these algorithms are a common source of deadlocks in the ThreadPool (which does not enable reentrant code). Finally, consider the usual case of a serial app that computes results in a certain order. When this applimsdnmagazine.com FREE Download With Full Support! Interactive diagram components www.nwoods.com December 2008 19 http://msdn.microsoft.com/magazine/cc817398 http://msdn.microsoft.com/concurrency http://msdn.microsoft.com/concurrency http://www.nwoods.com http://www.nwoods.com http://www.msdnmagazine.com

Table of Contents Feed for the Digital Edition of MSDN Magazine - December 2008

MSDN Magazine - December 2008 Contents Toolbox CLR Inside Out Advanced Basics Cutting Edge Patterns In Practice Team System Real-World WF Visual Studio OBA Tools SOA Data Access Geneva Framework Test Run Foundations Windows With C++ Going Places End Bracket

MSDN Magazine - December 2008

For optimal viewing of this digital publication, please enable JavaScript and then refresh the page. If you would like to try to load the digital publication without using Flash Player detection, please click here.