APU: 一种精确评估超线程处理器算力消耗程度的方法

伴随着云计算的发展,以及软件即服务(SaaS)、方法即服务(FaaS)等服务框架的提出,数据中心作为服务的提供商,面临着持续性的资源管理挑战:一方面需要保证服务质量(quality of service, QoS),另一方面又需要控制资源成本.为了在提升资源使用率的同时确保负载压力在可承受范围内波动,一种精确衡量当前算力消耗程度的方法成为关键性的研究问题.传统的评估指标CPU利用率,由于虚拟化技术的成熟以及并行技术的发展,无法应对资源竞争所产生的干扰,失去了评估精度.而当前数据中心的主流处理器基本都开启了超线程技术,这导致评估超线程处理器算力消耗程度的需求亟待解决.为了应对这一评估挑战,基于超线程机制的理解以及线程行为的建模,提出一种评估超线程处理器算力消耗的方法 APU.同时考虑到不同权限的用户能访问的系统层级不同,还提出了两种实现方案:一种基于硬件层支持的实现,以及一种基于操作系统层支持的实现. APU方法利用传统CPU利用率指标作为输入,没有其他维度的需求,免去了新监测工具的开发部署代价,也无需特殊硬件体系结构的支持,确保该方法的通用性和易用性.最后通过SPEC基准测试程序进一步...

APU: Method to Estimate Computing Power Consumption of Hyper-threading Processors

With the development of cloud computing and service architectures including software as a service (SaaS) and function as a service (FaaS), data centers, as the service provider, constantly face resource management. The quality of service (QoS) should be guaranteed, and the resource cost should be controlled. Therefore, a method to accurately measure computing power consumption becomes a key research issue for improving resource utilization and keeping the load pressure in the acceptable range. Due to mature virtualization technologies and developing parallel technologies, the traditional estimation metric CPU utilization fails to address interference caused by resource competition, thus leading to accuracy loss. However, the hyper-threading (HT) technology is employed as the main data center processor, which makes it urgent to estimate the computing power of HT processors. To address this estimation challenge, this study proposes the APU method to estimate the computing power consumption for HT processors based on the understanding of the HT running mechanism and thread behavior modeling. Considering that users with different authorities can access different system levels, two implementation schemes are put forward: one based on the hardware support and the other based on the operating system (OS). The proposed method adopts CPU utilization as the input without demands for other dimensions. Additionally, it reduces the development and deployment costs of new monitoring tools without the support of special hardware architectures, thereby making the method universal and easy to apply. Finally, SPEC benchmarks further prove the effectiveness of the method. The estimation errors of the three benchmarks are reduced from 20%, 50%, and 20% to less than 5%. For further proving the applicability, the APU method is leveraged to ByteDance clusters for showing its effects in case studies.