Web Testing for Reliability Improvement - Advances in Computers - page 211

ciated with 10,279 bytes. We used the standard two-hour gap [32] to identify user
sessions here.
The overall traffic at the KDE web site is significantly heavier than that for
SMU/SEAS, with a daily average of 3563 Mbytes, 455,005 hits, 24,656 users, 29,029
s1 sessions, and 104,490 s2 sessions. Two different variations of session count were
used in dealing with the KDE data: the same two-hour gap cut-off we used for the
SMU/SEAS web site (labeled s1), and the 15 minutes cut-off more appropriate for
dynamic pages (labeled s2) [39] .
No matter which workload measure is used, the daily workload shows several
apparent characteristics for both web sites, as follows:
Uneven distribution and variability : The distribution is highly uneven and varies
from day to day, as represented by the peaks and valleys in workload plots,
which conforms to previously observed traffic patterns [1,21,39] . Among the
four workload measures, daily bytes and daily hits show larger variability in
relative magnitudes than daily users or daily sessions. This result indicates that
although the number of users or user sessions may remain relatively stable,
some users may use the web much more intensively than others, resulting in
larger variations in detailed web site workload measurements over time. The
relative differences for KDE tends to be smaller than that for SMU/SEAS, likely
due to the heavier traffic by a larger and more diverse user population world-
wide.
A periodic pattern that synchronize with error profile , which is characterized by
weekdays normally associated with heavier workload than during the weekends.
This pattern seems to conform to the self-similar web traffic [12] . In addition,
this periodic pattern are correlated or synchronized with daily error profile. This
fact indicates that these workload measures are viable alternatives for web soft-
ware reliability evaluation, because of the direct relation between usage and
possible failures for the web site's source contents demonstrated in such syn-
chronized patterns.
A long term stability for the overall trend , which can be cross-validated by ex-
amining the trend over a longer period. All the workload measures traced over a
year for SMU/SEAS showed long-term stability. This is probably due to the sta-
ble enrollment for SMU/SEAS and web site stability where no major changes
were made to our web-based services over the observation period.
Of the four workload measures, hits, users, and sessions can be extracted from
access logs easily and consistently. However, byte counting was somewhat problem-
atic, because “byte transferred” field was missing not only for the error entries but
also for many other entries. Further investigation revealed that most of these missing
entries were associated with files or graphics already in the users cache (“file not