Campus Event Calendar
Event Entry
What and Who
Partial Replication in Large Networks
Nicolas Schiper
University of Lugano
SWS Colloquium
Nicolas Schiper holds a MSc in computer science from EPFL and recently
defended his PhD thesis which was supervised by prof. Fernando Pedone at
the University of Lugano in Switzerland. He is currently a post-doc in
the same institution.
His main research interests are distributed systems and fault-tolerance.
SWS, RG1
Expert Audience
English
Note: We use this to send email in the morning.
Date, Time and Location
Tuesday, 16 February 2010
14:00
60 Minutes
E1 5
5th floor
Saarbrücken
Abstract
The popularity of Internet applications such as e-banking, on-line stores, and
social networks has tremendously increased in the past years. As a
consequence, our daily lives depend on computers more and more each day.
These applications typically run on many machines spread across several
geographical sites, or groups, to provide low response times to clients. In
these complex systems, hardware and software failures are common.
Providing high availability without sacrificing performance is thus of a prime
importance.
In this talk, we present protocols to achieve high availability through data
replication. We proceed in two steps. We first devise fault-tolerant atomic
multicast algorithms that offer message ordering guarantees, and allow
messages to be addressed to any subset of groups. We then build a partial
replication protocol that relies on this multicast service. The four atomic
protocols presented differ in which properties they ensure, namely disastertolerance
and genuineness. With the former property, entire groups may
contain faulty computers; with the latter, to deliver a message m, protocols
only involve the groups addressed by m.
Performance at the multicast layer is obtained by minimizing the latency to
deliver multicast messages. At the replication layer, data is partitioned among
groups. This allows for more scalability than found in traditional replicated
systems since sites only handle a fraction of the workload. We show how
convoy effects can appear in partially replicated systems and propose
techniques to reduce these effects. Experimental evaluations compare the
proposed solutions.
Contact
Brigitta Hansen
0681 - 9325691
--email hidden
Video Broadcast
Yes
Kaiserslautern
G26
206
passcode not visible
logged in users only
Uwe Brahm, 02/18/2010 16:18
Brigitta Hansen, 02/11/2010 16:33 -- Created document.
Brigitta Hansen, 02/11/2010 16:33 -- Created document.