Document Type
Dissertation
Date of Award
5-31-2021
Degree Name
Doctor of Philosophy in Computing Sciences - (Ph.D.)
Department
Computer Science
First Advisor
Xiaoning Ding
Second Advisor
Cristian Borcea
Third Advisor
Jason T. L. Wang
Fourth Advisor
Jing Li
Fifth Advisor
Qing Gary Liu
Abstract
Multi-threading and resource sharing are pervasive and critical in clouds and data-centers. In order to ease management, save energy and improve resource utilization, multi-threaded applications from different tenants are often encapsulated in virtual machines (VMs) and consolidated on to the same servers. Unfortunately, despite much effort, it is still extremely challenging to maintain high quality of service (QoS) for multi-threaded applications of different tenants in clouds, and these applications often suffer severe performance degradation, poor scalability, unfair resource allocation, and so on.
The dissertation identifies the causes of the QoS problems and improves the QoS of multi-threaded execution with three approaches. First, the dissertation identifies that the I/O performance of an application can be significantly affected by its computation on VMs. Particularly, the I/O inactivity problem is caused when the computation workload has consumed the CPU time allocated to virtual CPUs (vCPUs), preventing the I/O workload on these vCPUs from producing I/O requests. This problem can greatly degrade I/O performance and cause fairness issues in I/O scheduling. Second, on modern simultaneous multi-threading (SMT) processors, existing CPU schedulers are ineffective to schedule I/O workloads for high I/O performance and efficiency. Third, due to frequent synchronization and communication, multi-threaded workloads are more vulnerable to the contention for CPU time in clouds. As a result, these workloads suffer severe performance degradation and interference.
The dissertation presents three systems, VMIGRATER, vSMT-IO, and JUPITER, with each addressing a distinct research problem. Extensive evaluations on diverse multi-threaded applications, including DBMS, web servers, AI workloads, Hadoop jobs, and so on, show that these systems can significantly improve the QoS of multi-threaded applications in clouds.
Recommended Citation
Jia, Weiwei, "Improving multi-threaded qos in clouds" (2021). Dissertations. 1727.
https://digitalcommons.njit.edu/dissertations/1727
