MSDN Magazine - December 2008 - (Page 17)

on more cores. However, the lines at the bottom (light blue and dark blue) corresponding to the bigger workloads indicate that the throughput is actually increasing when more cores are added, while the throughput for the experimental empty workload shows a slight decrease. Additionally, larger workloads can experience more lock contention. As you can imagine, addressing all these pitfalls in a single, general-purpose optimization technique is difficult. siderations for Parallel Programming” at msdn.microsoft.com/magazine/ cc872852 and “CLR Inside Out: Using Concurrency for Scalability” at msdn.microsoft.com/magazine/cc163552.) Dangerous Tuning To make things more challenging, there are cases where the introduction of some logic designed to provide some performance improvement can have the opposite effect. A typical example is when such an improvement is designed to deliver general-purpose optimizations; for some applications it just adds a lot of overhead, which in turn results in performance degradation. Picture a case where various synchronization mechanisms are added to an app to function in multiple cores (to distribute work fairly, for instance), doing the synchronization adds complexity, which may result in unnecessary overhead for simple apps where an easier solution may have worked better. In highly concurrent applications, it is hard to identify significant tuning parameters. For example, maximizing the CPU usage does not necessarily result in throughput improvement; while in single-processor machines it does, in manycore machines it is sometimes better to not maximize the CPU usage and, instead, make threads “waste” some time. Memory Hierarchy Awareness Taking into account that memory hierarchy is one of the most important issues for performance tuning—exploiting memory locality, especially for higher numbers of cores, is indispensable. In non-uniform memory access (NUMA) architectures, accessing memory locations in “distant” nodes is expensive and can hurt performance significantly. Understanding the underlying memory hierarchy is important for scheduling purposes. For example, a way of keeping memory accesses more localized in a multiprocessor environment would be to maintain threads working on a particular task in the same processor. The ThreadPool is not NUMA-aware; however, the addition of strategies such as work stealing inherently takes advantage of memory locality. Similar tasks tend to remain in the same node due to local queuing and dequeuing. In general, remember that the efficient utilization of a cache can result in significant performance improvements. (For more information on memory hierarchy, see “Paradigm Shift: Design Con- Noise Noise is yet another factor that makes performance tuning difficult. All the experiments presented in this discussion were run on Advanced Document Imaging Toolkits Zero-footprint Viewing, 3 Months of Free Gold Support, No Sweat. Features include Runtime Royalty-Free OCR, PDF annotations, ISIS support, CAD/CAM and WPF support, AJAX web image viewing with no plug-ins, and more—Take document and photographic imaging to the .NExT step. Visit Atalasoft.com for web demos, video tutorials, and a free trial. Atalasoft, Your .NET Imaging Partner Call us toll free at 866-568-0129 msdnmagazine.com www.atalasoft.com December 2008 17 http://msdn.microsoft.com/magazine/cc872852 http://msdn.microsoft.com/magazine/cc872852 http://msdn.microsoft.com/magazine/cc163552 http://www.atalasoft.com http://www.Atalasoft.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.