Efficiency Scheduling Policies Of Hybrid Virtualized CPU-GPU Resource

CPU(Central Processor Unit) is good at dealing with complex computational logic and processes, such as artificial intelligence, physics simulation. In contrast, GPU(Graphic Processing Unit) is good at computing big data due to have many cores. All in all, CPU has fewer but relatively powerful cores, while GPU has a large number of common cores. Currently the system consists of a multi-core CPU and a many cores GPU has become a powerful platform for processing hybrid workload.However, how to share CPU and GPU resources among multiple virtual machines in the cloud data center remains a problem. First, it will cause confusion performance and waste of resources of GPU intensive workload due to the lack of effective GPU resource scheduling policy. Second, CPU intensive workloads will take a lot of CPU and IO resources due to the heterogeneity of the workloads. This will affect the performance of GPU intensive workloads. Moreover their SLA(Service Level Agreement) requirements are difficult to be guaranteed. Finally, due to cloud data centers need to maintain a lot of network and user connections, limited network bandwidth resources has become a major bottleneck affecting the user experience.We propose a distributed architecture based on virtualization, which is a lightweight framework that efficiently allocating mixed GPU and CPU resources. The framework consists of two parts, wherein the first portion is run in a single physical node responsible for GPU, CPU and network bandwidth resource scheduling. It uses Microsoft Hook(Hook) technology, by intercepting the associated API(Application Programming Interface) for each virtual machine. The second part of the framework is an independent master machine, accept incoming user requests and automatically make arbitration.Because GPU resource scheduling problem is mainly the work of previous studies, resource scheduling in this article will only involve in GPU resource scheduling when it is necessary. In order to provide more choice, this framework incorporates several resource scheduling policy, which consisted of one GPU resource scheduling policy from previous work, two new CPU resource scheduling policies and two network bandwidth resource scheduling policies. Respectively they are Open-loop control policy, Adaptive control policy, Fairness-allocation policy and Sharing-allocation policy. Both of the first two scheduling policies are for CPU, while the others are for network bandwidth resources. Scheduling policies above are implemented by the automatic control technology, and enhance the stability of the load runtime performance and robustness, effective control of Qo S level mixed load.