A Case Study in Natural Language Based Web Search - Natural Language Processing and Text Mining - page 78

hensive and authoritativeonline destinationforthose in need ofboth reliable back-
ground information and breaking news ...GlobalSecurity.org's unique positioning
enables it to reachboth a targeted and large diversified audience. Thecontent of
the website is updatedhourly, asevents around the worlddevelop, providing in-
depth coverage of complicatedissues. The breadth and depth ofinformation onthe
siteensures a loyal repeat audience. This issupplementedbyGlobalSecurity.org's
unique visibility in the mass media, which drives additionalgrowth”[9]. The direc-
tor ofGlobalSecurity.org, John Pike, regularlyprovides commentary and analysison
space and security issues to PBS, CNN, MSNBC,Fox, ABC, CBS,NBC, BBC,NPR,
and numerous print and online publications. Inpowering thissite, InFact serves the
information searchneedsofawell-establisheduser community of100,000, consist-
ingofnews reporters, concerned citizens, subject matter experts, senior leaders, and
junior staff and interns.
5.4.2 Operational Considerations
Whenpreparing the GlobalSecurity.org deployment,oneof our primeconcerns was
the response timeofthesystem.Forthis reason, we kept the data sizeofthe partition
indices small enough so that most operationsoccur in memory and diskaccess
is minimal. Wesplit the GlobalSecurity.org data across two index chunks, each
containing roughly14GBof data in each partitionindex. Another concern was
havingsu cient capacity to handle the user load. To account for future user tra c,
wespecifiedthe deployment for 2-3times the maximum expected load ofabout
11,000 queries perday. Thisleft us with twocloned partition indices perindex
chunk.In addition, we wantedahotbackupoftheentiresite, in case ofany
hardware failures, and tosupport useachtime we are rollingout newfeatures.
Fig. 5.4. The GlobalSecurity.org home page.
Anotherarea of concern was the distribution ofquery types. Our systemhas
significantly varying average response time and throughput (measuredinqueries/
minute) dependingonthe typeofqueries beingexecuted. We assumed that users