多应用负载性能干扰预测与隔离相关问题研究

Cloud datacenters usually adopt workload co-location strategy, which means place and execute multiple workloads in one single physical computing unit. Thus the physical computing resources can be shared by allocating logical computing resources to the workloads, which achieves the goals of increasing the datacenter resource utilization and reducing the infrastructural and operational cost. However, when the computing tasks grow, the competition among workloads for the shared computing resources will become intense, and the interference among workloads will lead to the low system performance and the uncertainty of execution time. This proposal aims at addressing the low resource utilization problems and the QoS degradation related issues caused by cross-workload interference while multiple workloads running together in datacenter, and will research the cross workload interference quantization and prediction methods, as well as the corresponding interference reduction countermeasures. The detailed research approaches include: the modeling of multi-applications interference and the proactive detection on cross workload interference, the interference-aware system parameter tuning algorithm, and the operating system level performance interference isolation mechanism. The approaches together will provide a theoretical and technical basis to increase the datacenter resource utilization and reduce the infrastructure cost and energy consumption.

云数据中心采用负载整合的方式，即将多个应用负载放置在一个物理计算单元中运行，通过为应用负载分配逻辑计算资源的方式共享物理计算资源，以期达到提高数据中心的资源利用率以及降低投资与运行成本的目的。当计算任务增多时，多应用负载对共享计算资源的竞争更加激烈，所产生的性能干扰会导致各应用的性能下降以及负载执行时间的不确定性。本项目针对数据中心多应用负载共同运行时因性能干扰而导致资源利用率低下与应用负载服务质量下降等问题，研究多应用负载间性能干扰的量化与预测方法，以及降低性能干扰的措施与方法。内容包括：多应用性能干扰模型建立以及应用间性能干扰的前摄性检测、性能干扰感知的参数配置调优、系统级性能干扰隔离机制等，为提高数据中心资源利用率，降低成本与能耗提供理论与技术支撑。

在现代互联网系统中，软件系统除了要能够完成某种特定功能外，还需要快速响应用户请求和不间断提供在线服务。随着云计算技术的发展，越来越多的IT和互联网企业都开始提供云计算服务并将自身业务向云端转移，云计算的基础设施（即云数据中心）的数量和规模也在迅速的增长。在云环境中，应用对共享资源的竞争将会引起性能干扰，性能干扰导致数据中心资源利用率低，进而增加成本投入。本课题围绕数据中心性能干扰问题展开研究，目标是降低性能干扰并提高资源利用率，研究内容包括：多应用性能干扰预测、干扰感知的自适应参数调优、性能干扰隔离及资源优化。