Fair Scheduling On Dynamic Heterogeneous Chip Multiprocessors

Compared to the Symmetric Chip Multi-core Processors(SCMP), Asymmetric Chip Multi-core Processors(ACMP) configure different types of cores,the large cores have a complicated structure, high power consumption and powerful performance, such as the superscalar processors which can develop the instruction-level parallelism of serial program. At the same time, The small cores have simple structure, low power consumption, small area which can develop thread-level parallelism of parallel program. In fact, ACMP will efficiently improve system performance and reduce power consumption only when the granularity of cores matches the characteristics of tasks.Dynamic Heterogeneous Chip Multiprocessors,(DHCMP) comprise lots of small basic cores that can be integrated into different type of larger logical cores. DHCMP can meet the needs of different tasks and improve system performance and power consumption.However, DHCMP only provides the ability of reconfiguring logical cores.whether DHCMP accurately perceives behavior characteristics and resource requirements of tasks,reasonably allocates resource and fairly progresses tasks depends on task scheduling and resource allocation algorithm. Based on We design new task scheduling and resource allocation algorithm to improve the overall performance, service quality and energy efficiency.The main research work and achievements in this paper can be include into the following two aspects:1. We design a resource allocation algorithm:PCPRA. In each operating system scheduling interval, resource allocator first distribute same logical cores for each task. After set a time, sort tasks according to their performance. Find out the task that need least resource, reduce the granularity of its logical core, and increase the granularity of the logical core on which the task need most resources.2. After analyzing the deficiency of traditional task scheduling algorithm on DHCMP, we design the performance-aware fair task scheduling algorithm PDP and PCFS according to the new definition of the fairness of task scheduling. These two algorithms can dynamically perceive task performance to measure resources requires of tasks, allocate logical CPU time. So,they can effectively improve the service quality,system performance and energy efficiency. In the experiment, we simulate the SCMP, ACMP, DHCMP platform through the TFlex simulator. First, we implement CFS, DWRR, DP scheduling algorithm on SCMP platform, ADWRR scheduling algorithm on ACMP platform, DP, CFS, EDP, PDP, PCFS scheduling algorithm on DHCMP platform(using the PERA resource allocation algorithm). The experimental result shows that, PDP, PCFS scheduling algorithms have advantage of scheduling fairness, system performance and efficiency. Then, we implement the PCPRA/PCFS combination, which is better than the PERA/PCFS group on system performance and not inferior to the PERA/PCFS. It indicates that PCPRA is better than PERA.