The Commonwealth Server Project

The phenomenal growth of the WWW is imposing considerable strain on Internet resources and Web servers, prompting numerous concerns about the Web's continued viability. The success of high-performance Web servers in alleviating these performance problems is ultimately limited, unless Web services are designed to be inherently scalable. The Commonwealth project is designing, implementing, and evaluating a prototype architecture and a set of associated protocols for scalable Web services. The Commonwealth architecture for hosting scalable Web services allows scalability through (1) parallel processing on tightly-coupled nodes within a Web site, and (2) load distribution across loosely-coupled Web sites. Commonwealth's underlying philosophy is to achieve a WEALTH of performance through the use of COMMON components, and to do so along an incremental upgrade path.

The Commonwealth project is pursuing basic research in four main areas:

• Data placement (caching, prefetching, and dissemination protocols);
• Resource management (scheduling, load balancing, and admission control protocols);
• Networking (routing, packet redirection, and connection migration) and
• Middleware Services (replication and resource discovery).
These problems are being addressed with particular attention to the issue of high workload variability, which---in previous work by the Oceans Group at Boston University---was shown to pose significant problems for the design of Internet-based systems.

Project Co-Leaders:


Azer Bestavros
Mark Crovella

Project Sponsors:


The National Science Foundation
Hewlett Packard Laboratories
Microsoft Research

Summary

At the lowest layer, cluster-based servers must employ some method for directing incoming service requests to individual hosts (i.e., connection routing). Previous methods for connection routing have typically employed a centralized node which handles all incoming requests. In contrast, the Commonwealth project has developed a distributed approach, called Distributed Packet Rewriting (DPR), in which all hosts of the distributed system participate in connection routing. We argue that this approach promises better scalability and fault-tolerance than the centralized approach. Four variants of DPR have been implemented and we have compared their performance. We find that DPR provides performance comparable to centralized alternatives, measured in terms of throughput and delay under the SPECweb96 benchmark. We believe that DPR is particularly attractive both for small scale systems and for systems following the emerging trend toward increasingly intelligent I/O subsystems.

Resources

• Distributed Packet Rewriting and its Application to Scalable Server Architectures, Technical Report BUCS-TR-98-003 (Compressed postscript).
• Online DPR Measurement demonstration page
  Only functions for machines on network 128.197.10

Participants

• Luis Aversa
• Scott Harrison
• Jun Liu
• David Martin

Summary

This work is studying the problem of task assignment in distributed systems with highly variable task sizes. Work to date has focused on task assignment policies that make use of knowledge of incoming task sizes, which is a capability we are designing into the Commonwealth Server.

Our initial focus has been on Size Interval policies: each host is assigned tasks from a certain size interval. We have studied two variants of Size Interval Task Assignment (SITA): policies that assign equal load to each host (SITA-E) and policies that assign unbalanced load to different hosts (SITA-V).

We have shown that under SITA-E, FCFS queueing typically outperforms the processor-sharing discipline by a factor of about two, as measured by mean waiting time and mean slowdown (waiting time of task divided by its service time). We have compared the FCFS/SITA-E policy to the processor-sharing case analytically; in addition we have compared it to a number of other policies in simulation. We find that the benefits of SITA-E are present even in small-scale distributed systems (four or more hosts).

SITA-V purposely operates the server hosts at different loads, and directs smaller tasks to the lighter-loaded hosts. The result is that SITA-V provably decreases the mean task slowdown by significant factors (up to 1000 or more) where the more heavy-tailed the workload, the greater the improvement factor. We have evaluated the tradeoff between improvement in slowdown and increase in waiting time in a system using SITA-V, and found conditions under which SITA-V represents a particularly appealing policy.

Resources

• M. Harchol-Balter, M. E. Crovella, and C. D. Murta, "To Queue or Not to Queue: When Queueing is Better Than Timesharing in a Distributed System," BUCS-TR-1997-017, October 31, 1997; in postscript.
• M. E. Crovella, M. Harchol-Balter, and C. D. Murta, "Task Assignment in a Distributed System: Improving Performance by Unbalancing Load," BUCS-TR-1997-018, October 31, 1997; in postscript.

Participants

• Mor Harchol-Balter
• Cristina Duarte Murta

Summary

This work is applying a number of observations of Web server usage to create a realistic Web workload generation tool which mimics a set of real users accessing a server. The tool, called Surge (Scalable URL Reference Generator) generates references matching empirical measurements of 1) server file size distribution; 2) request size distribution; 3) relative file popularity; 4) embedded file references; 5) temporal locality of reference; and 6) idle periods of individual users. We have reviewed the essential elements required in the generation of a representative Web workload, and addressed the technical challenges to satisfying this large set of simultaneous constraints on the properties of the reference stream. We have found significant evidence that Surge exercises servers in a manner significantly different from other Web server benchmarks.

Resources

• P. Barford and M. E. Crovella, "Generating Representative Web Workloads for Network and Server Performance Evaluation," BUCS-TR-1997-006, November 4, 1997; in postscript.

Participants

• Paul Barford

• Meeting Schedule and Assigned Readings
• Fall 1996 CS-835: Seminar on Networked Computing Systems
• The Oceans Group
• Commercial Products of Interest to Commonwealth
• Pointers of Interest

Created on: 1996.08.27
Updated on: 1998.11.30