Performance Profiling Method And Tool Based On PMU For NUMA Platform On VM Environment

Because of its scalability and easy of programming, the NUMA(Non-Uniform Memory Access) platform has become the main infrastructure for high performance computing and cloud computing. At the same time, due to its highly isolated to the underlying hardware, more acceptable to users, compatible to myriad OSes and applications, low risk and easier to maintain, the virtual machine technology is used widely both in personal computers and commercial servers. NUMA hardware platform with virtual machines has become a main system environment for cloud computing.There are many performance profiling tools can be used on NUMA platform on non virtual machine environment, and also have many research works about performance profiling and optimization of NUMA, but recent research papers and reports shows that there are still a lot of work about performance profiling and optimization on NUMA platform we can do in virtual machine environment. However, due to the guest OS cannot be aware of the underlying NUMA hardware and lack of performance profiling tool, the research work about performance analysis and optimization on NUMA platform in virtual machine environment relatively less, and focused on management of VMs and vcpu scheduling algorithm.This paper aims at conclude the main performance factors of NUMA platform, design and implements a NUMA platform performance profiler which adapt to the virtual machine(we called it VMMprof temporarily), and proposes a kind of method which can find and evaluate the performance bottlenecks of NUMA in virtual machine using VMMprof from the viewpoint of micro architecture. Therefore it can guide the vcpu scheduling and optimizing performance of the systems or applications.This paper firstly analyzed the characteristics of the micro architecture and memory accessing of the NUMA platform, and concluded the main performance factors, which contains the traditional factors(data locality, LLC contention, Sharing overhead and IMC contention) and the factor contention of GQ trackers which we proposed firstly. Then, we verified this five factors through a benchmark program we designed.After we have learned the main performance factors of NUMA platform, we go about the implementation of VMMprof, which based on the PMU in the processor. VMMprof perfectly and firstly support the performance profiling of NUMA Uncore part in xen, and also support self defined profile model, remote operation, multiple users and has a remarkably user-friendly interface. And then, we has carried out the verification of its correctness.Finally, we creatively proposes a kind of method which can find and evaluate the performance bottleneck factors of NUMA in virtual machine using VMMprof. The experimental results show that we can effectively find and evaluate the node’s performance bottlenecks of NUMA platform in virtual machine environment through this method. And then we put forward some optimization advices for thoses bottlenecks from the perspective of vcpu scheduling.