Academic Open Internet Journal

www.acadjournal.com

Volume 6, 2006

A Web Cache On A Fast Network

Jan Haluza, Ivan Dolezal

1. Introduction

WWW caches are set up inparticular to increase WWW throughput and to decrease webpage access time(the time between creation of a request and its satisfaction).
An improvement of networktransfer rates, as well as fast connectivity of terminals (users’ computers),could, however, bring increasingly embarrassing questions to an administrator,such as "Is the WWW proxy cache still an asset? What is raison d’etre forits deployment in a situation when there is a sufficient bandwidth bothin
LAN/MAN and for internationalbackbone connectivity?"

2. Experiment Description

A simple measurement wasdone to simulate (with wget client) behavior of a user accessing some webpageswith her browser. The frequency – a function of occurences of data transferrates of unique connections from server to user in one of fifty intervalsfrom 100 B/s to 10 MB/s with logarithmical division – was the followed-upquantity.

The measurement was donein one of localities connected to backbone of a national research and educationnetwork CESNET. WWW proxy cache was made of Squid running on Linux (PCIntel III/500 MHz, 384 MB RAM, Fast Ethernet). WWW client wget was runningon a PC of the same HW configuration.

The measurements were doneduring normal cache operation whilst processing approx. 45 requests persecond.

A) Perfectly Random SelectionOf Webpages By User (Perfect Cache Miss)

An extreme case is expectedthat none of client URL requests repeats, implying that none of them canbe replied directly by the WWW cache. Thus, the WWW proxy cache is a redundantelement in the process of requesting and sending a web page. All the repliesmust travel from WWW servers. The goal is to recognize the influence ofWWW proxy cache on transfer rate from WWW servers to user over direct transferfrom WWW servers to user. A set of pseudorandom URLs not contained in cache’sdisk space was generated for the purpose of this experiment.

The algorithmof pseudorandom selection of URLs:
1. Create a set ofrandom words
Repeat a pseudorandom selectionof words from a dictionary (for example words containing two letters withtheir defined distance).
2. Create a fileof URLs for the given set of words
Submit the words to a publicsearch engine to find webpages containing these words.
3. Remove invalidURLs
Check availability of thewebpages.

B) Perfect URL Selection(Perfect Cache Hit)

Flip side of the previousexperiment is a situation, where every client’s request can be answeredby the cache. In this experiment the WWWcache is initially prefetched withobjects that will later be requested by the client. The goal is to comparethe frequency of data transfer rates through WWW cache with direct accessto the URLs. The results are presented in graph below (section D) as 100%proxy.

C) WWW Servers In NetworksConnected To Fast Backbone

In this case a set of webservers located in metropolitan area networks connected to the fast nationalbackbone was requested. Comparability of obtained values is limited bya number of available WWW servers (due to small number of connected academicalinstitutions).

D) The Reality

In order to check the realeffect of the WWW cache between clients and WWW servers the sets of URLspartially present in the WWW cache are requested. It is a fact that thehit byte rate (an indicator of general effectiveness) of transparent WWWproxy caches of CESNET is approx. 40 %. Therefore this value was takenas a criterion for this experiment.

This measurement was donewith a set of URLs taken from a real traffic of the WWW proxy cache farm.According to the fact that effectiveness of DNS cache running for the WWWproxy is 97 %, this is the represen-tative sample of really requested servers.A main web page was requested from each of these servers. It is to be assumedthat locations of these WWW pages on the servers couldn’t essentially changetheir transmission rates. Also, using the cached DNS records eliminateseasily multiple occurences of clients’ requests of unique URLs from theset.

The combination of objectsthat were present/absent in the WWW cache, corresponds with 40 % cache’shit byte rate.

3 Results

•  Even if a client is connectedto a fast backbone directly via Fast Ethernet interface, the average datatransfer speed from server to client is approx. 0.5 MB/s. Since the URLset used in the experiment results mainly in oversea queries, we can expectthe same results in the similar academic networks in Europe.
• Transactions made with WWW cachedo not have some radical negative impact on operations missed by the cache,even for the clients connected to the fast backbone.
• If the cache’s hit byte rateis approx. 40 %, the average speed of transfer of a WWW document from asource to a user is almost three times higher.

Even if a WWWcache is connectedto a fast backbone network, the availability of WWWdocuments is even morecut back by network conditions all the way from WWW server to browser,as well as by efficiency of a WWW source. Thus WWW cache is still an assetfor fast backbones.

5 Reference

[1] A. Feldmann, R. Caceres,F. Douglis, G. Glass, M. Rabinovich, Performance of Web proxy caching inheterogenous bandwidth environments, Proceedings of IEEE INFOCOM'99, March1999

[2] SPEC Group, SPECweb99benchmark, Technical Report, 1999 (available from http://www.spec.org/osg/web99/)

[3] C. Maltzahn, K. Richardson,D. Grundwald, Performance issues of enterprise level Web proxies, in: Proc.1997 ACM Sigmetrics Int. Conf. on Measurement and Modelling of ComputerSystems, June 1997, pp.13-23

[4] G. Banga, P. Druschel,Measuring the capacity of a Web server, in: Proc.1997 USENIX Symp. on InternetTechnology and Systems, December 1997

[5] B.M. Duska, D. Marwood,M.J. Feeley, The measured access characteristics of World-Wide-Web clientproxy caches, in: Proc. USENIX Symp. on Internet Technology and Systems,December 1997