Characterization of containers in emerging applications: Microservices, FAAS and GPUS

Darbaz, Haldun Umur

Permalink

https://hdl.handle.net/2142/102816 Copy

Description

Title

Characterization of containers in emerging applications: Microservices, FAAS and GPUS

Author(s)

Darbaz, Haldun Umur

Issue Date

2018-11-29

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

Kim, Nam Sung

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Docker
• Containers
• Microservices
• FaaS
• GPUs
• TLB
• Computer
• Architecture
• Operating
• Systems
• Page
• Tables
• Virtualization
• Linux

Abstract

Containers have enabled new computing paradigms such as Functions-as-a- Service in data centers today. Containers are inherently more lightweight than virtual machines. This is caused by the fact that containers share the kernel with the host system, removing the need for a two-dimensional page walk. Containers also do not require a hypervisor. They rely on thin management layers in container frameworks and existing Linux functionality. Linux process and resource management features such as cgroups and namespaces are tightly integrated to containers. This allows for simple management and isolation of containerized applications. Docker is currently the most prominent container framework. This thesis utilizes Docker containers to create data center use cases with databases, web servers, graph analytics, Functions-as-a-Service, and GPU-accelerated stencil, lower-upper decomposition, object tracking and neural network applications. Furthermore, this thesis analyzes Docker Engine performance by bringing up containers and breaks down bring-up overheads at function granularity. The virtual memory management aspects of Docker containers are also characterized with a focus on container infrastructure, page tables and page faults. This thesis reports on average 59.86% duplicated page table entries and 35.7% duplicated page faults across four containerized processes sharing a core. Additionally, this thesis identifies the source of 40% of container bring-up overhead and attributes it to memory allocation, garbage collection and process creation in Go and Linux. This thesis also identifies a 7% slowdown in containerized GPU applications with NVIDIA-Docker compared to native execution. Finally, this thesis provides guidance to architects for enabling container support in high-performance architectures, and identifies future work to be done in the area.

Graduation Semester

2018-12

Type of Resource

text

Permalink

http://hdl.handle.net/2142/102816

Copyright and License Information

Copyright 2018 Haldun Umur Darbaz

Owning Collections