[Colloquium] Ahsan Pervaiz Dissertation Defense/Apr 25, 2023

[Megan Woodward]

This is an announcement of Ahsan Pervaiz's Dissertation Defense.
===============================================
Candidate: Ahsan Pervaiz

Date: Tuesday, April 25, 2023

Time:  11:00 am CST

Remote Location: https://uchicago.zoom.us/j/95848322579?pwd=Mi9ObEkyUTAvZmJIR2pRQ0U2SXp1UT09

Location: JCL 390

Title: Frameworks for General-Purpose Adaptation in Computing Systems

Abstract: Modern computer systems are becoming increasingly complex. Apart from functional correctness, they are expected to provide strict guarantees on their quality-of-service, expressed as goals over important quantifiable metrics such as latency, in the face of unpredictable changes in operating conditions and workloads during the execution of the system. Furthermore, these systems expose a plethora of tunable parameters that, combined with the unpredictable external conditions, impact the quality-of-service of the system. It is a well-established fact that no single configuration of the tunable parameters is optimal for all workloads and operating conditions, hence, systems are required to dynamically adapt their tunable parameters to cope with dynamic changes in workloads and operating conditions.

The systems community views adaptation as a crucial capability for systems to deliver reliable quantitative behavior in the presence of dynamic external factors. However, developing modules for robust adaptation is difficult and requires specialized knowledge of machine learning and/or control theory. This increases the burden on developers who are expected to be experts in the aforementioned fields along with their system specific domain. To alleviate this burden, prior work suggests packaging adaptation modules as a library or in the runtime of a language. During execution, systems can simply instantiate these modules by providing them with the goal that needs to be met and the parameters that can be adapted to meet the goal. Once instantiated, the frameworks then continually monitor the relevant behavior of the system and adapt the configurations on behalf of the system to ensure that the system continues to meet its goal.

However, A major limitation of prior frameworks is that they are implemented for a specific, narrow set of goals and knobs. Hence, they cannot be used for complex adaptive systems that must meet different goals using different sets of knobs for different deployments, or different execution stages of one deployment. For such scenarios developers are expected to embed different frameworks in their systems to support different goals. This, in turn, increases the size of the system code base and makes development and deployment more difficult.

In our research we explore the benefits of providing a single generalized adaptation framework that is agnostic of knobs and goals. We show the deployment and runtime benefits of using such a framework in a number of real-world systems including a networked video analytics pipeline.
Our research shows that it is not only possible to implement a generalized adaptation framework but that using such a framework is more favorable from a development, deployment and performance perspective.

Another problem in the same domain is that of colocation. To maximize resource utilization and efficiency, system administrators, such as cloud providers, colocate multiple systems on the same hardware. However, when multiple adaptive systems are colocated they negatively interfere with each other leading to misbehavior which results in significant degradation of quality-of-service.

Prior works have recognized the problem of negative interference and have suggested many approaches to mitigate it. Such approaches require all colocatable systems to be enumerated beforehand. However, this restricts the system administrators' ability to colocate different systems. Furthermore, they impose severe restrictions on the systems that use them. For example, they impose restrictions on the mehanisms that the systems can use for adaptation. Similarly, they require information of internal details of the colocatable systems to be shared amongst each other. Often times prior work also requires all colocated systems to delegate their adaptation to a monolithic adaptation module. Such restrictions render approaches suggested by prior work impractical for real-world use where proprietary information can seldom be shared, and development can seldom be coordinated between stakeholders.

In the latter part of this body of work we explore a general framework that can be used by all systems individually to allow harmonious execution of colocated adaptive systems without any restrictions on the methods used for adaptation, explicit coordination or information sharing. We show that such a system is beneficial because it provides a degree of freedom to stakeholders to develop their systems independently without worrying about the other systems that they may be colocated with their systems.

Hence, the contributions of this body of work are frameworks for: (1) adding complex, generalized adaptation in computing systems and (2) ensuring their successful and harmonious execution when they need to be colocated with each other.

Advisors: Henry Hoffmann

Committee Members: Henry Hoffmann, Shan Lu, Haryadi Gunawi





-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20230424/61e9c613/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: ahsan_phd_thesis.pdf
Type: application/pdf
Size: 8166310 bytes
Desc: ahsan_phd_thesis.pdf
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20230424/61e9c613/attachment-0001.pdf>