高利用率驱动多核QoS保障关键技术研究

Multicore system largely improves the system performance, by allowing multiple tasks to run at the same time. However, because of the lacking of guarantees to QoS goals of co-running tasks, the use of multicore becomes a hazard. On one hand, increasing the number of tasks risk the violation of their QoS goals. On the other hand, although reducing the number of tasks running in the system ensures the good performance, the cost is expensive since large amount of resources are not used. Such problem becomes the constraints on many important computing platforms, such as the cloud computing, and it draws extensive attentions and research efforts..The scientific problem we attack in this project is to provide QoS guarantee when the multicore system is under high load. We propose a novel mechanism to research on it. The first contribution we aim to achieve is to deliver a proper measurement of the multicore system load which should accurately quantify the computing power used by current workload and the remaining capacity. Secondly, we further explore the scenario when the use of the resources is saturated to enrich the measurement, and thus to quantify the range of the high load. Thirdly, within the range of the high load, we propose techniques to guarantee the QoS goals of tasks, and to optimize the performance and resource utilization of the system. Our overall goal is to provide sound techniques to increase the number of task capacity in the multicore system with the QoS goal guaranteed, and hence to increase the resources utilization in the system.

对高质量计算服务的需求，使得多核系统的利用面临着两难的处境：利用多核架构特征并发执行更多任务面临着QoS目标不可控的风险；而降低任务数量提高任务性能又导致了大量资源闲置的问题。本项目所研究的科学问题即，当多核系统处于高负载时如何保障任务QoS目标的问题。先前的研究表明，通过优化资源分配提高任务性能及系统输出可行。基于此，本项目的研究思路如下：1.多核系统在多个任务并发执行时，我们从资源利用的角度研究对其负载能力进行精确测量和表达的方法；2.特别是，在资源被过度使用的情况下对系统负载能力进行衡量，并由此确定系统高负载能力的上限；3.在界定的高负载区间内，研究多维度且分配特征丰富的资源分配方法来达到任务QoS目标可控的理想状态。研究的目的是通过对多核系统可承受高负载区间的精确把握，实现区间内任务QoS的可控，进而为并发任务数量的有效增加和资源的高效利用提供充分的科学依据。

当前的数据中心面临着一系列问题，例如多任务时资源利用率极低，此时如何保证QoS目标。我们又该如何去衡量量化能耗模型，进而研究如何提高资源利用率。项目通过研究阿里巴巴集群数据集，发现其策略具备“弹性”和“塑性”，从而保证其QoS，我们并总结了其半容器的优点与代价。项目中，我们提出了智能能耗计算方法（Sensible Energy Accounting, SEA），并对DRAM做了相关的测试，即预测计算执行任务时DRAM的能耗(SEnsible DRAM Energy Accounting, SEDEA)。我们的SEDEA能耗预测模型预测错误均值约为6.5%，标准偏差小于13%。与其他预测模型ES、 PTA、 DReAM，SEDEA的预测相比，误差最小。SEDEA还可以用于探索闪存分区和内存分区之间的协同关系，节省的系统能耗（实验中节省了8.7%）。此外，针对多核系统中，以任务QoS为目标的应用，我们开发Chameleon系统架构。项目中，我们还完成了GPGPU内核的多层次的表征与优化，及Spark性能调优的研究。