a provider-side view of web search response time
DESCRIPTION
A Provider-side View of Web Search Response Time. Yingying Chen, Ratul Mahajan, Baskar Sridharan , Zhi -Li Zhang (Univ. of Minnesota) Microsoft. Web services are the dominant way to find and access information. Web service latency is critical to service providers as well. revenue -20%. - PowerPoint PPT PresentationTRANSCRIPT
A Provider-side View ofWeb Search Response Time
YINGYING CHEN, RATUL MAHAJAN, BASKAR SRIDHARAN, ZHI-LI ZHANG (UNIV. OF MINNESOTA)
MICROSOFT
Web services are the dominant way to find and access information
Web service latency is critical to service providers as well
Bing
revenue-20%
Latency+2 sec
revenue-4.3%
Latency+0.5 sec
Understanding SRT behavior is challenging
t
300+tS
RT
(ms)
M T W Th F S Su
peak off-peak
200+t
t
SR
T (m
s)
Our work
Explaining systemic SRT variation
Identify SRT anomalies
Root cause localization
Client- and server-side instrumentation
HTML header
Brand header
BoP scriptsQuery results
Embedded images
query
π ππ π ππ
π hπππ
π πππππ
π hπππ‘π π1
π πππ π»πππΏ
π π΅ππ
π hπππ‘π π2
π πππππ
π πππ
π π πππππ‘
π π π
π π‘π
on-load
Referenced content
Impact Factors of SRT
π ππ
network browser queryserver
π hππππ ππππππ hπππ‘π π1π πππ π»πππΏπ π΅πππ hπππ‘π π2π ππππ π πππππ‘π πππ‘π π ππ πππ πππππ π π‘π
Primary factors of SRT variation
Apply Analysis of Variance (ANOVA) on the time intervals
ΖSRT
varianceVariance explained by time interval k
Unexplainedvariance
Primary factors: network characteristics, browser speed, query type Server-side processing time has a relatively small impact
network browser queryserver
π hππππ πππ π»πππΏπ π΅πππ πππ π π πππππ‘π πππ‘ π π ππ ππ π π‘π
Expl
aine
d va
rianc
e (%
) 60
40
20
0
Variation in network characteristics
RTT
Explaining network variations
Residential networks send a higher fraction of queries during off-peak hours than peak hours
Residential networks are slower
residential enterprise
RTT
(ms)
25%1.25t
t
Residential networks are slowerResidential networks send a higher fraction of queries during off-peak hours than peak hours
residential unknownenterprise
Variation in query type
Impact of query on SRT Server processing timeRichness of response page
Measure: number of image
Explaining query type variationPeak hours Off-peak hours
Browser variations Two most popular browsers: X(35%), Y(40%) Browser-Y sends a higher fraction of queries during off-peak hours Browser-Y has better performance
Browser-X Browser-Y
Javascript exec time
82%1.82t
t
Summarizing systemic SRT variation Server: Little impact
Network: Poorer during off-peak hours
Query: Richer during off-peak hours
Browser: Faster during off-peak hours
Detecting anomalous SRT variations
Challenge: interference from systemic variations
Week-over-Week (WoW) approach
+ Seasonality + Noise
Comparison with approaches that do not account for systemic variations
WoW One Gaussian model of
SRT
Change point
detection
False negative 10% 35% 40%
False positive 7% 17% 19%
Conclusions
Understanding SRT is challengingChanges in user demographics lead to systemic
variations in SRT
Debugging SRT is challenging Must factor out systemic variations
ImplicationsPerformance monitoring
Should understand performance-equivalent classes
Performance managementShould consider the impact of network, browser, and
query
Performance debugging End-to-end measures are tainted by user behavior
changes
Questions?