Intel® 64 and IA-32 Architectures Optimization Reference Manual

MULTICORE AND HYPER-THREADING TECHNOLOGYIf the control flow of a multi-threaded application contains a workload in which only50% can be executed in parallel, the maximum performance gain using two physicalprocessors is only 33%, compared to using a single processor. Using four processorscan deliver no more than a 60% speed-up over a single processor. Thus, it is critical tomaximize the portion of control flow that can take advantage of parallelism. Improperimplementation of thread synchronization can significantly increase the proportion ofserial control flow and further reduce the application’s performance scaling.In addition to maximizing the parallelism of control flows, interaction betweenthreads in the form of thread synchronization and imbalance of task scheduling canalso impact overall processor scaling significantly.Excessive cache misses are one cause of poor performance scaling. In a multithreadedexecution environment, they can occur from:• Aliased stack accesses by different threads in the same process• Thread contentions resulting in cache line evictions• False-sharing of cache lines between different processorsTechniques that address each of these situations (and many other areas) aredescribed in sections in this chapter.8.1.2 Multitasking EnvironmentHardware multithreading capabilities in Intel 64 and IA-32 processors can exploittask-level parallelism when a workload consists of several single-threaded applicationsand these applications are scheduled to run concurrently under an MP-awareoperating system. In this environment, hardware multithreading capabilities candeliver higher throughput for the workload, although the relative performance of asingle task (in terms of time of completion relative to the same task when in a singlethreadedenvironment) will vary, depending on how much shared executionresources and memory are utilized.For development purposes, several popular operating systems (for exampleMicrosoft Windows* XP Professional and Home, Linux* distributions using kernel2.4.19 or later 2 ) include OS kernel code that can manage the task scheduling and thebalancing of shared execution resources within each physical processor to maximizethe throughput.Because applications run independently under a multitasking environment, threadsynchronization issues are less likely to limit the scaling of throughput. This isbecause the control flow of the workload is likely to be 100% parallel 3 (if no interprocessorcommunication is taking place and if there are no system bus constraints).2. This code is included in Red Hat* Linux Enterprise AS 2.1.3. A software tool that attempts to measure the throughput of a multitasking workload is likely tointroduce control flows that are not parallel. Thread synchronization issues must be consideredas an integral part of its performance measuring methodology.8-3