Performance Optimization Research For MIC-CPU Symmetric Mode

Recent years, supercomputers have acquired continuous development and already become an important manifestation of national science and technology strength in the era of information. However, limited by the physical extremity of transistor and the current standard of technology manufacturing constraints, the development of CPU has slowed down and thus cannot meet people's pursuits. In this case, a viable approach is to continue to increase computing cluster size. But the increased size will inevitably lead to a linear growth of the entire system power consumption and it's not acceptable. So beginning in 2007, people started using GPGPU to calculate high-density floating-point tasks to assist CPU as accelerator or be known as co-processor. We call systems consisting of different types of computing components heterogeneous computing system. As the most famous processor vendor, INTEL released an advanced many-core accelerator MIC in2012, which has been more widely used so far.MIC emplys the same x86 instruction sets with common CPU, running a streamlined Linux system we called ?OS. Therefore, their share similar programming models.MIC's computing core can be used as stand-alone computing nodes involved in cluster calculation and share relative position with CPU logically. We call this programming mode Symmetric Mode. Processes running on CPU and MIC communicates with each other through MPI(Message-Passing Interface). As common MPI applications running on CPU-only clusters, the communication brings about assignable expenses.This article focuses on MIC-CPU Symmetric Mode, while evaluating application executing expenses under this mode and making a study of correlating optimization research. The achievements are mainly:1. Under MIC-CPU symmetric mode, computing units consist of CPU and MIC which gears into CPU through PCI-e port. Each computing node in a cluster connects with each other through high-speed network. Because two different processes can be located on any computing units,communication in symmetric mode is complex.This article evaluates the performance of communications and programs in MICCPU mode, including communication in a single node or in two different nodes.Meanwhile, the article takes a study of how communication performance affect application performance of NPB and Linpack;2. Applications under MIC-CPU symmetric mode can actually be regarded as MPI application. In this part of work, we first make native application optimization on MIC, using SIMD and Threads expand optimization methods. Then load balance optimization is taken among CPUs and MICs, to avoid performance loss due to load inbalance;3. Due to communication performance difference between computing units of symmetric mode and communication demand difference of processes of an application,the article try to get satisfied performance promote via rational allocating processes to different symmetric components. In detail, the article abstracts this problem as a QAP problem and use ant algorithm to solve it;...