Fine-Grained I/O Fairness Analysis
in Virtualized Environments
Sewoog Kim
Dept. of Computer Science
Dankook University Yongin, Korea, 448-701
Dongwoo Kang
Dept. of Computer Science
Dankook University Yongin, Korea, 448-701
Jongmoo Choi
Dept. of Computer Science
Dankook University Yongin, Korea, 448-701
ksu1110@dankook.ac.kr ABSTRACT
rediori@dankook.ac.kr choijm@dankook.ac.kr
A virtualized system consists of a couple of independent and isolated software components such as hypervisor, control domain, and virtual machines, which leads I/O requests to be processed across several protection domains. These complicated inter-domain interactions frequently results in I/O performance degradation and fairness violation. In this paper, we propose a new virtualization-aware fine-grained I/O analysis framework and observe I/O behavior in a multimodal and multidimensional viewpoint. Our observations reveal that fairness is not preserved at various layers due to the combination of diverse causes including I/O scheduling and CPU scheduling. Also, some mechanisms that worked well in a non-virtualized system do not match well with a virtualized system, which contributes the unfairness. The observations trigger us to devise a simple fairness enforce technique that exploits the reservation and limit concepts to balance the I/O bandwidth shares among virtual machines. Real implementation based experimental results have shown that our proposal can enhance I/O fairness without considerable overheads.
consolidation and low TCO (total cost of ownership). To reap out these advantages, the virtualization technology is being studied and employed actively in both academia [1, 2, 3] and industry [4, 5, 6].
The virtualization technology introduces a new software layer called hypervisor or VMM (Virtual Machine Monitor) that governs the behavior of virtual machines [7]. The hypervisor can be classified into two types, bare-metal based [1, 4] and host based [8, 9, 10], according to the layer where it locates. Also, according to whether the hypervisor allows the modification of an operating system (a.k.a. GuestOS) in a virtual machine, it can be categorized into two types, paravirtualization [2] and full virtualization [4].
The goal of the hypervisor is managing various logical resources efficiently and fairly among virtual machines. To achieve the goal, it makes use of a variety of policies. For instance, the borrowed virtual time scheduling [1] and proportional-share algorithm [11] are used for CPU virtualization, while the difference engine [12] and ballooning [4] used for memory virtualization. Also, for I/O virtualization, the hypervisor makes use of the I/O emulation technique [8], IOMMU [13] and IDD (Isolated Device Domain) [14]. The IDD, which is commonly used in a XEN-based virtualized system, is an isolated and privileged control domain that takes fully charge of the I/O device management.
The coexistence of isolated and independent software components makes I/O requests being handled across several protection domains. For instance, a I/O request issued by an application is delivered to the related I/O device through the GuestOS, hypervisor, and control domain using various complex data structures such as I/O ring and event channel [1]. Also, a response is returned with the similar interactions among the protection domains. These interpositions of isolated domains in the I/O path often lead to performance degradation and fairness violence in virtualized environments [16].
To explore these issues more quantitatively and with a system-wide viewpoint, we propose a virtualization-aware fine-grained I/O analysis framework. It consists of two modules, on-line monitor and off-line analyzer. The on-line monitor has two features, fine-grained and multimodal feature. It defines five focal layers, namely I/O request, I/O queuing, I/O dispatch, I/O completion and I/O response layer to examine I/O fairness in a fine-grained fashion. In addition, it keeps track of not only I/O events but also scheduling information. The off-line analyzer has a multidimensional feature that shows the monitoring results with various aspects such as timeline or each GuestOS’s share. Using
Categories and Subject Descriptors
D.4.8 [Operating System]: Performance – Operational Analysis; D.4.4 [Operating System]: Communication Management – Input/output; D.2.2 [Software Engineering]: Design Tools and Techniques – Modules and interfaces
General Terms
Measurement, Experimentation
Keywords
Virtualization, I/O fairness, Fine-Grained Analysis, Multimodal, Multidimensional
1. INTRODUCTION
Virtualization is a technology that abstracts physical resources into multiple logical ones so that multiple virtual machines can run at the same time on a physical machine. It supports various advantages such as high utilization, isolation, live migration,
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee.
RACS’12, October 23-26, 2012, San Antonio, TX, USA. Copyright 2012 ACM 978-1-4503-1492-3/12/10…$15.00
热门文档