PerfIso: Performance Isolatin forCommercial Latency-Sensitive Services

Posted 2021-08-15 银灯玉箫

我的论文笔记框架（Markdown格式）

PerfIso: Performance Isolatin forCommercial Latency-Sensitive Services
{2018}, {Iorgulescu}, {ATC}

Summary
写完笔记之后最后填，概述文章的内容，以后查阅笔记的时候先看这一段。注：写文章summary切记需要通过自己的思考，用自己的语言描述。忌讳直接Ctrl + c原文。

Research Objective(s)
The average load is far lower than the peak load used for provisioning, leading to resource under-utilization. PerforISo is used to colocate batch jobs with production latency-sensitive services on over 90,0000 serviers, increasing average CPU utilization from 21% to 66% for off-peak load without impacting tail latency.

off-peak
美
英
adj.非高峰期的；淡季的

Background / Problem Statement
We refer to the latency-sensitive user-facing services as primary tenants. All resources of a machine need to be available for them, since they generate the revenue that pays for the actual machines. Thus, the primary always runs unrestricted and unmodified.
tenant
美 ['tenənt]
英 ['tenənt]
n.房客；租户；佃户

Primary tenants comprise the latency-sensitive services which are governed by strict response-time SLOs. They are characterized by the following:

1. a complex layered architecture – hard to model or predict, sicne response are computed in parallel and then aggregated.
2. short tail latnecy – any layer can severely impact query response times.
3. high bursty nature – the service frequently spawns a large number of workers in a short period of time (order of microseconds).
   bursty
   美 ['bɜrsti]
   英 ['bɜː®sti]
   adj.突发式数据的；集送数据的；脉冲数据的；阵发性的

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？
We propose a new technique called CPU Blind Isolation, which restricts which cores secondary tennants use based on core utilization information read from the OS.

The key idea is to ensure that the primary always has some headroom(i.e., buffer idle cores), to absorb any primary worker-threads that wake up.

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。
————————————————
版权声明：本文为CSDN博主「银灯玉箫」的原创文章，遵循CC 4.0 BY-SA版权协议，转载请附上原文出处链接及本声明。
原文链接：https://blog.csdn.net/lilele12211104/article/details/118719875