{ATORI : Efficient and Fair Resource Partitioning by Sacrificing Short-Term Benefits for Long-Term}(文章标题)
{2021}, {Rohan Basu Roy}, {ISCA}
{引用格式}
Summary
SATORI is a novel strategy to partition multi-core architectural resources to achieve two conflicting goals simultaneously: increasing system throuphout and achieving fairness among the co-located jobs.
Research Objective(s)
作者的研究目标是什么?
Background / Problem Statement
研究的背景以及问题陈述:作者需要解决的问题是什么?
Co-locating workloads on chip multiprocessors(CMPs) is an attractive approach for cloud computing service providers as it enables them to improve resource utilization and lower the capital and opertional costs of runing their dataceters.
Adding more shared resources(e.g. memory bandwidth) makes the problem even more challenging, as previous works such as CoPart, Heracles, an PARTIES have demonstrated,. These works have taken two different approaches to optimize the partitioning of multiple shared resources. In the first approach, partitioning of individual resource is done separately, but the decisions are communicated to reach a global optimal resource partition.
For example, CoPart maintains two seperate finite state machines(FSM), one for shared chaed and one for memory bandwidth. These FSMs are not joint or linked but are aware of each other’s decisions. In the second approach, partitioning of individual resources is performed in one dimension at a time (similar to a gradient descent method). For example, Heracles and PARTIES perform resource partitioning in a gradient descent sytle where partitioning of one resource is explored first before adjusting the allocations for other resources. CoPart is simple, but may not scale well to multiple resources, unlike PARTIES that can scale well to multiple resources.
Method(s)
作者解决问题的方法/算法是什么?是否基于前人的方法?基于了哪些?
SATORI develops a new approach grounded in Bayesian Optimization(BO) theory to intelligently explore multi-resource partitioning configuration space. In particular, SATORI’s BO based approache enables us to design a pratically-feasible technique that can work on a real system in an online fashion. The key intuition behind using BO is to build simple and just-accurate-enough models for finding near-optimal solutions, without requiring offline profining, instrumentation, offline deep learning based training which may incur high-overhead opon the nature of the offline traning dataset. As our evaluation confirms, tolerating a slight inaccuracy of the model allows SATORI to achieve near-optimal configurations in an online fashion. SATOP enables fast and efficient navigation of large configuration space to find the optimal resoruce partition configuration. It performs joint exlporation of mutiple resources simultaneously – removing the limitation of existing approaches that maintain one finite state machine for each shared resource or explore one resource dimension at a time.
Evaluation
作者如何评估自己的方法?实验的setup是什么样的?感兴趣实验数据和结果有哪些?有没有问题或者可以借鉴的地方?
Conclusion
作者给出了哪些结论?哪些是strong conclusions, 哪些又是weak的conclusions(即作者并没有通过实验提供evidence,只在discussion中提到;或实验的数据并没有给出充分的evidence)?
Notes
(optional) 不在以上列表中,但需要特别记录的笔记。
References
(optional) 列出相关性高的文献,以便之后可以继续track下去。