Understanding Resource Usage and Performance in Wide-Area Distributed Systems

Abstract

Many Internet services employ wide-area frameworks to deliver

exponentially growing network traffic to end users with low response

time. These systems typically leverage a large number of remote nodes

at the edge of the Internet, which makes the systems difficult to

develop and test. Therefore, federated testbeds are essential

infrastructures for developing wide-area systems because they allow

researchers to deploy new services under realistic network conditions.

In this dissertation, we study resource usage in PlanetLab to

understand and characterize user behavior in federated testbeds. We

also present Lsync, a low-latency file transfer system for

coordinating remote nodes in wide-area platforms, including testbeds.

To support the development of new network services on a global scale,

the next generation of federated testbeds are under active

development, but very little is known about resource usage in these

shared infrastructures. We conduct an extensive study of the usage

profiles in PlanetLab that we collected for six years by running

CoMon, a PlanetLab monitoring service. We examine various aspects of

node-level behavior as well as experiment-centric behavior, and

describe their implications for resource management in federated

testbeds. We find that the usage is much different from shared

compute clusters, that conventional wisdom does not hold for

PlanetLab, and that several properties of PlanetLab as a network

testbed are largely responsible for this difference.

We also present a low-latency file transfer system, Lsync, that can be

used as a synchronization building block for wide-area distributed

systems where latency matters. While many distributed systems depend

on fast data synchronization for coordinating remote nodes, current

data dissemination systems focus on efficiency for open client

populations, rather than focusing on completion latency for a known

set of nodes. In examining this problem, we find that optimizing for

latency produces strategies radically different from existing

distribution tools, and can dramatically reduce latency across a wide

range of scenarios. Lsync performs novel node selection, scheduling,

and adaptive policy switching that dynamically chooses the best

synchronization method using information available at runtime. Our

evaluation results show that Lsync reduces latency by more than a

factor of 14 compared to a widely used synchronization tool, and makes

most remote nodes fully synchronized even under frequent file updates.