Managing Contention for Shared Resources on Multicore Processors - Case Study Example

Managing Contention for Shared Resources on Multicore Processors Introduction The processing speed of acomputer system is dependent on the processors and the memory capacity of a system. The need to enhance the speed of computer systems is what led to the invention of the concept of parallel computing. A parallel computer is a set of more than one processor core that is able to work together simultaneously to solve a computational problem. Parallel computing is used in powerful computers, some with hundreds or thousands of processors. The idea of parallel computing has led to the invention of modern powerful computer systems such as supercomputers that are able to multifunction. The advantages of parallel computing include saving on the time required for computing and providing concurrency where the multiple processors can solve several computational problems simultaneously. The modern computer systems with multicore processors apply the concept of parallel computing in their operation. In parallel computing, the computer systems share several hardware resources such as LLC and memory controllers to enhance their operation. Several processors cores are assigned to a common memory resource. Processors operating under the same memory resource may compete for the shared memory resource thus, causing traffic and congestions. This is referred to as contention. Contention causes slowing down of the computer system thus reducing the performance of the system (Yuejian, 2012). There are two types of contentions namely; communication contention and memory contention. Communication contention occurs when several processor cores contend for a common communication link. This causes traffic and performance degradation, and in turn slows performance. Memory contention on the other hand, occurs when several processors compete for resources from the same memory module. In a test to demonstrate how contention for the shared resources affects the operation of the computer system, three applications namely Soplex, Sphinx, and Namd were run simultaneously on an Intel quad core xeon system. Soplex, Sphinx, and Namd were paired to run in the same memory domain in different schedules. The result of the combinations of applications indicated a dramatic difference between the different pairs of applications. The applications run as a whole performed 20 percent better with the best schedule, while by running Soplex and Sphinx applications simultaneously the performance was great as 50 percent. The Soplex and Sphinx pair of combination sharing the same memory module was considered as the best schedule. Techniques for reducing contention The results of the test indicate that to run a contention, free environment applications should be assigned according to the best possible schedule, although a contention-oblivious scheduler might accidentally occur on the best schedule and cause it to run as the worst schedule. To prevent any case of contentions, a contention aware scheduler should be better positioned to select an application combination that performs better. The other method that can be used to reduce contention in the computer system is through proper organization of the memory modules. Here, the main memory of the system is partitioned into local and global memory. The local memory is attached to one processor while the global memory is only accessible by sending request through the interconnection network. The local memory is further partitioned into private and shared memory. The shared memory is accessible by all processors while the private, memory is only accessible by the processor to which it is attached. The private memory helps reduce network traffic, communication and memory contention. Memory allocation data structures should be effectively allocated to the shared memory modules. The local memory modules easily store codes and data structures private to the processors. This reduces the chances of communication and memory contention. Contention can, therefore, be reduced by having several memory modules attached to the interconnection network with a common access throughput that is equal to the number of shared memory modules (Fountain, 2006). The future solution to reducing cases of memory and communication contention is through development of a modern operating system that is programmed with contention-aware scheduling techniques. The operating system will be designed to run on multicore systems that apply the concept of parallel computing. Cache contention This type of contention occurs when more than two threads are assigned to run on cores of the same memory module.The cache consists of lines that are allocated to hold the memory of the threads as they issue cache requests. When the thread requests a line that cannot be located on the in the cache because the cache is full, there occurs a cache miss therefore anew cache line must be allocated. The cases of many occurrences of cache misses’ causes cache contention. Contention occurs when the cache line needs to be allocated but the cache is full; hence, some of the data must be displaced in order to create a new line for the newly introduced piece of data. (Zwick, 2011). Techniques for reducing cache contention The solution to avoidance of cache contention is to create a scheduler that is programmed to avoid contention. However, there is need to determine a way of predicting the occurrence of contentions. Cache contentions are mostly caused by cases of cache miss. The cases of contentions can be determined by predicting the miss rate of the cache. Prediction can be done by determining the number of times per instruction the thread fails to find the line in the cache. If the thread issues lots of cache misses, it has a large working set thus, it suffers contention. The other solution to cache contention is by avoiding the reuse of cache data. If a thread does not reuse its cached data, for example, when video streaming, it will not be affected by contention regardless of the amount of data it brings to the cache. The reason behind this is because a thread requires very little space to keep in the cache, the data that it actively uses. This method requires creation of models for shared cache based on memory reuse (Louise, 1998). Power management and distribution is important in managing contention within the computer system. This can be done by turning off the unused cores or by clustering. Clustering involves assigning the workload to a few available servers while turning off the rest. Clustering may, however, lead to lowering the performance of the system due to competing of the shared resources. A long term solution to cache contention is the development of a modern operating system that is scheduled to avoid cache misses thus reducing the high miss rate and contention. The scheduler integrated within the operating system will be designed with a power distribution intensity that will be responsible for management of power utilization and distribution without affecting the performance of the system. References Fountain T. J. (2006). Parallel Computing: Principles and Practice. Washington: Cambridge University Press. Louise J.T. (1998). Run Time Adaptive Cache Management. Illinois: University of Illinois Press. Yuejian X. (2012). Efficient Shared Cache Management for Multicore Processors. Washington: BiblioBazaar. Zwick M. J. (2011). Predicting Cache Contention in Multicore Processor. Washington: Cambridge University Press. Read More