CATCloud: closing semantic gaps of CPU interfaces for precise autoscaling in the cloud

Sampat, Pratik Rajesh

Permalink

https://hdl.handle.net/2142/124222 Copy

Description

Title

CATCloud: closing semantic gaps of CPU interfaces for precise autoscaling in the cloud

Author(s)

Sampat, Pratik Rajesh

Issue Date

2024-04-25

Director of Research (if dissertation) or Advisor (if thesis)

• Ghose, Saugata
• Xu, Tianyin

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Operating System
• CPU allocation
• Autoscaling
• Cloud computing
• Resource Management

Abstract

Precise CPU allocation for a multi-programmed computer is crucial to application performance and resource efficiency, but is notoriously difficult under dynamic cloud workloads, where multiple users executing diverse applications often share the CPUs. We argue that the fundamental problem is rooted in the mismatch of the existing CPU allocation interface between the cloud and the OS---while the cloud represents CPU resources as a percentage quota of the host CPU (i.e., millicpu), the OS interprets CPU resources as time-shared quota slices allowed to run within a defined period. The cloud interface's disregard for periodicity stems from the fundamental difficulty of capturing fine-grained application runtime behavior in userspace. Consequently, existing solutions rely on coarse-grained, surrogate metrics such as CPU utilization, throttle, and queue lengths, leading to slow and imprecise allocation. We present CATCloud, an OS extension that closes the semantic gap of cloud CPU allocation. CATCloud views CPU resources as a shared bandwidth interface and implements a millisecond-scale CPU bandwidth autotuner for quota and periodicity. Implemented in the OS scheduler, CATCloud realizes observability of fine-grained run time and yield time behavior of target applications; which was previously opaque to the userspace autoscalers. By continuously capturing historical data, it accurately estimates the short-term CPU period and quota requirements. With an execution latency of only a few milliseconds, CATCloud can quickly and effectively react to bursty, dynamic workloads with simple statistical algorithms. We show that CATCloud significantly outperforms state-of-the-art techniques in terms of responsiveness, precision, and efficiency. Our evaluation on various cloud workloads shows that CATCloud can improve CPU efficiency by on an average of 27.8%, up to 81.3% and performance improvements on average of 27.91%, up to 152.5% with negligible memory and compute overheads, over existing autoscaling solutions

Graduation Semester

2024-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/124222

Copyright and License Information

Copyright 2024 Pratik Rajesh Sampat

Owning Collections