Cloud Control with Distributed Rate Limiting
Barath Raghavan, Kashi Vishwanath, Sriram Ramabhadran,
Kenneth Yocum, and Alex C. Snoeren
Department of Computer Science and Engineering
University of California, San Diego
ABSTRACT
Today’s cloud-based services integrate globally distributed re-
sources into seamless computing platforms. Provisioning and ac-
counting for the resource usage of these Internet-scale applications
presents a challenging technical problem. This paper presents the
design and implementation of distributed rate limiters, which work
together to enforce a global rate limit across traffic aggregates at
multiple sites, enabling the coordinated policing of a cloud-based
service’s network traffic. Our abstraction not only enforces a global
limit, but also ensures that congestion-responsive transport-layer
flows behave as if they traversed a single, shared limiter. We
present two designs—one general purpose, and one optimized for
TCP—that allow service operators to explicitly trade off between
communication costs and system accuracy, efficiency, and scalabil-
ity. Both designs are capable of rate limiting thousands of flows
with negligible overhead (less than 3% in the tested configuration).
We demonstrate that our TCP-centric design is scalable to hundreds
of nodes while robust to both loss and communication delay, mak-
ing it practical for deployment in nationwide service providers.
Categories and Subject Descriptors
C.2.3 [Computer Communication Networks]: Network manage-
ment
General Terms
Algorithms, Management, Performance
Keywords
Rate Limiting, Token Bucket, CDN
1. INTRODUCTION
Yesterday’s version of distributed computing was a self-
contained, co-located server farm. Today, applications are increas-
ingly deployed on third-party resources hosted across the Inter-
net. Indeed, the rapid spread of open protocols and standards
like Web 2.0 has fueled an explosion of compound services that
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, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
SIGCOMM’07, August 27–31, 2007, Kyoto, Japan.
Copyright 2007 ACM 978-1-59593-713-1/07/0008 ...$5.00.
script together third-party components to deliver a sophisticated
service [27, 29]. These specialized services are just the beginning:
flagship consumer and enterprise applications are increasingly be-
ing delivered in the software-as-a-service model [9]. For example,
Google Documents, Groove Office, and Windows Live are early ex-
amples of desktop applications provided in a hosted environment,
and represent the beginning of a much larger trend.
Aside from the functionality and management benefits Web-
based services afford the end user, hosted platforms present sig-
nificant benefits to the service provider as well. Rather than de-
ploy each component of a multi-tiered application within a partic-
ular data center, so-called “cloud-based services” can transparently
leverage widely distributed computing infrastructures. Google’s
service, for example, reportedly runs on nearly half-a-million
servers distributed around the world [8]. Potential world-wide scale
need not be limited to a few large corporations, however. Recent
offerings like Amazon’s Elastic Compute Cloud (EC2) promise to
provide practically infinite resources to anyone willing to pay [3].
One of the key barriers to moving traditional applications to the
cloud, however, is the loss of cost control [17]. In the cloud-based
services model, cost recovery is typically accomplished through
metered pricing. Indeed, Amazon’s EC2 charges incrementally
per gigabyte of traffic consumed [3]. Experience has shown, how-
ever, that ISP customers prefer flat fees to usage-based pricing [30].
Similarly, at a corporate level, IT expenditures are generally man-
aged as fixed-cost overheads, not incremental expenses [17]. A
flat-fee model requires the ability for a provider to limit consump-
tion to control costs. Limiting global resource consumption in a
distributed environment, however, presents a significant technical
challenge. Ideally, resource providers would not require services
to specify the resource demands of each distributed component a
priori; such fine-grained measurement and modeling can be chal-
lenging for rapidly evolving services. Instead, they should provide
a fixed price for an aggregate, global usage, and allow services to
consume resources dynamically across various locations, subject to
the specified aggregate limit.
In this paper, we focus on a specific instance of this problem:
controlling the aggregate network bandwidth used by a cloud-based
service, or distributed rate limiting (DRL). Our goal is to allow a set
of distributed traffic rate limiters to collaborate to subject a class of
network traffic (for example, the traffic of a particular cloud-based
service) to a single, aggregate global limit. While traffic policing is
common in data centers and widespread in today’s networks, such
limiters typically enforce policy independently at each location [1].
For example, a resource provider with 10 hosting centers may wish
to limit the total amount of traffic it carries for a particular service
to 100 Mbps. Its current options are to either limit the service to
100 Mbps at each hosting center (running the risk that they may all
- 1
- 2
- 3
- 4
- 5
前往页