Transcript Web and Intranet Performance: a quantitative analysis

Part III
Web and Intranet
Performance Issues
(Book, chap. 4)
8 Menascé & Almeida. All Rights Reserved.
1
Learning Objectives
•Present server architecture and
performance issues
•Discuss perception of performance
•Introduce Web infrastructure components
•Discuss Web server workload
•Examine bandwidth, latency, and traffic in
the Web
•Introduce capacity planning questions
8 Menascé & Almeida. All Rights Reserved.
2
Web Server Performance
Problems
Unpredictable nature of information
retrieval and service request over the
World-Wide web
• load spikes: 8 to 10 greater than avg.
• high variability of document sizes: from
103 to 107 bytes
8 Menascé & Almeida. All Rights Reserved.
3
Web Server Elements
HTTP
server
TCP/IP
O.S.
hardware
8 Menascé & Almeida. All Rights Reserved.
Contents:
. HTML
. graphics
. audio
. video
. other
4
HTTP request
• Method (GET, HEAD, PUT, POST)
• URI (Universal Resource Indicator),
actually URL (a sub-set of URI)
• Protocol version
• Additional information (optional): types
of documents accepted, authentication,
etc.
8 Menascé & Almeida. All Rights Reserved.
5
HTTP response
• Status line
• Metainformation (description about
information sent)
• Information (document file, or output
from server-side (servlet) application)
8 Menascé & Almeida. All Rights Reserved.
6
Combination of HTTP and TCP/IP
• HTTP defines a request-response
interaction;
• HTTP is a ``stateless’’ protocol;
• one connection per object;
• TCP connection setup overhead;
• mandatory delays due to the protocols;
• small Web objects and the TCP ``slow
start’’ algorithm
8 Menascé & Almeida. All Rights Reserved.
7
HTTP request-response steps
• map the server to an IP address (DNS);
• establish a TCP/IP connection with the
server;
• transmit the request (URL,method,etc);
• receive the response (HTML text or
other information);
• close the TCP/IP connection.
8 Menascé & Almeida. All Rights Reserved.
8
HTTP 1.0 interaction
0 RTT
syn
TCP conn.
syn
1 RTT
ack
client sends
HTTP req.
Server residence
time
dat
ack
dat
2 RTT
client parses
HTML doc.
syn
syn
Server residence
time
3 RTT
client sends
req. for image
dat
4 RTT
dat
image begins
to arrive
8 Menascé & Almeida. All Rights Reserved.
9
HTTP 1.1 interaction
0 RTT
syn
TCP conn.
syn
1 RTT
ack
client sends
HTTP req
dat
ack
2 RTT
client parses
HTML doc.
client sends
req. for image
dat
Server residence
time
ack
dat
3 RTT
ack
Server residence
time
dat
image begins
to arrive
8 Menascé & Almeida. All Rights Reserved.
10
HTTP 1.0 and 1.1 interaction
0 RTT
0 RTT
syn
TCP conn.
TCP conn.
syn
1 RTT
ack
client sends
HTTP req.
dat
ack
2 RTT
client parses
HTML doc.
syn
3 RTT
client sends
req. for image
Server
time
syn
syn
1 RTT
client sends
HTTP req
ack
Server
time
client parses
HTML doc.
client sends
req. for image
ack
dat
ack
Server
time
dat
ack
3 RTT
dat
image begins
to arrive
HTTP 1.0
8 Menascé & Almeida. All Rights Reserved.
image begins
to arrive
Server
time
dat
2 RTT
dat
dat
4 RTT
syn
dat
HTTP 1.1
11
Hardware, O.S.
•
•
•
•
•
Processors (speed, number)
Memory (capacity)
Disks (speed, capacity)
NIC: bandwidth
OS reliability, performance, scalability,
robustness; TCP/IP implementation!
8 Menascé & Almeida. All Rights Reserved.
12
Contents
• Contents can heavily impact on
performance:
• web site popularity
• size of user community
(extranet - hard to say; intranet - can be
determined/controlled)
• overall size and type of objects
(MM files are 100 to 10,000 times
bigger as HTML documents)
8 Menascé & Almeida. All Rights Reserved.
13
Where are the delays?
• Browser
– R’browser
• Network
– R’network
• Server
– R’server
• User response time: Rr
– Rr = R’browser + R’network + R’server
– Rr = R’cache
8 Menascé & Almeida. All Rights Reserved.
or
14
Anatomy of an HTTP transaction
End user
Client Browser
Network
Server
click
R’C
R’B
Data returned
from cache
HTTP Request
R’
N1
Rr
R’s
R’N2
Data
Display
8 Menascé & Almeida. All Rights Reserved.
Server
residence
time
Average Response Time
• Usually R’cache << R’network + R’server
• pc denotes the fraction of times the data are
found in the local cache
• R’cache: response time when the data are
found in a local cache
Exp(Rr) = pc x R’cache + (1-pc) x Rr
8 Menascé & Almeida. All Rights Reserved.
16
Impact of the Browser’s Cache
• 20% of the requests are serviced by the
local cache
• local cache response time = 400 msec
• average response time for remote Web
sites = 3 seconds
R = pc x R’cache + (1-pc) x Rr
R = 0.20x 0.4 + (1-0.20) x 3.0
R = 2.48 sec
8 Menascé & Almeida. All Rights Reserved.
17
Impact of the Browser’s Cache
• What if we increase the size of the local
cache?
• Previous experiments show that tripling the
cache size would raise the hit ratio to 45%.
Thus,
R = pc x R’cache + (1-pc) x Rr
R = 0.45x 0.4 + (1-0.45) x 3.0
R = 1.83 sec
8 Menascé & Almeida. All Rights Reserved.
18
Bottlenecks
• As the number of clients and servers
grow, overall performance is
constrained by the performance of
some components along the path from
the client to the server.
• The components that limit system
performance are called bottlenecks
8 Menascé & Almeida. All Rights Reserved.
19
Example of a Bottleneck
• A home user is unhappy with access times to
Internet services. To cut response time down,
the user is considering replacing the
processor of his/her desktop with one twice
as fast.
What will be the response time improvement if I
upgrade the speed of my desktop computer?
8 Menascé & Almeida. All Rights Reserved.
20
Example of a Bottleneck (cont’d)
for an average page:
• avg. network residence time:
• 7,500 msec
• avg. server residence time:
• 3,600 msec
• avg browser time:
• 300 msec
• Rr = R’browser + R’network + R’server = 300+7,500+3,600
• Rr = 11,400 msec = 11.4 sec
8 Menascé & Almeida. All Rights Reserved.
21
Example of a Bottleneck (cont’d)
• Percentage of time:
%x = R’x / (R’browser + R’network + R’server )
• browser = 300/11,400 = 2.14 %
• network = 7,500/11,400 = 65.79 %
• server = 3,600/11,400 = 31.57 %
8 Menascé & Almeida. All Rights Reserved.
22
Example of a Bottleneck (cont.)
• The CPU upgrade affects mainly the browser
time:
• R’Nbrowser ~ 1/2 x R’browser = 1/2 x 300 = 150
msec
• RNr = R’Nbrowser + R’network + R’server
• RNr = 150 + 7,500 + 3,600 = 11.25 sec.
• Therefore if the speed of the PC were doubled, the
response time would decrease only by
Rr/RNr = 11.40/11.25 = 1.3%
8 Menascé & Almeida. All Rights Reserved.
23
WEB workload is hard to
forecast
• (WEB performance depends heavily on
workload)
• Number of WEB clients: 107 and rising
• Randomness of visits
• large diversity in products (browsers,
servers, cache/proxies)
• unpredictable network delays, due to
netw. bandwidth and congestion
8 Menascé & Almeida. All Rights Reserved.
24
Perception of Performance
• WWW user:
• fast response time
• no connection refused
• Web administrators:
• high throughput
• high availability
Need for quantitative measurements
8 Menascé & Almeida. All Rights Reserved.
25
WWW Performance Metrics (I)
• connections/second
• Mbits/second
• response time
• user side
• server side
• errors/second
8 Menascé & Almeida. All Rights Reserved.
26
WWW Performance Metrics (II)
Web site activity indicators
• Visit: a series of consecutive Web page requests
from a visitor within a given period of time.
• Hit: any connection to a Web site, including in-line
requests, and errors.
• Request = successful hit.
• Metrics
• hits/day
• visits/day
• unique visitors/day
• pages
views/day
8 Menascé
& Almeida.
All Rights Reserved.
27
WWW Performance Metrics (III)
Web Advertising Measurements
• Exposure metrics (visits/day, pages/day)
• site exposure
• page exposure
• banner exposure
• Interactivity metrics
• visit duration time
• inter-visit duration
• visit depth (total # of pages a visitor is exposed to
during a single visit to a Web site)
8 Menascé & Almeida. All Rights Reserved.
28
Example of Performance Metrics
The Web site of a travel agency was
monitored for 30 minutes and 9,000 HTTP
requests were counted. We want to assess
the server throughput.
• 3 types of Web objects
– HTML pages: 30% and avg. size of 11,200 bytes
– images: 65% and avg. size of 17,200 bytes
– video clips: 5% and avg. size of 439,000 bytes
8 Menascé & Almeida. All Rights Reserved.
29
Example of Performance Metrics
Throughput
• in terms of requests:
– (No. of requests)/(period of time) =
– 9,000/(30 x 60) = 5 requests/sec
• In terms of bits/sec per class
– (total requests x class % x avg. size) /
(period of time)
8 Menascé & Almeida. All Rights Reserved.
30
Example of Performance Metrics
• HTML throughput (Kbps)
• 9,000 x 0.30 x (11,200 x 8) / 1,800 = 131.25
• Image throughput (Kbps)
• 9,000 x 0.65 x (17,200 x 8) / 1,800 = 436.72
• Video throughput (Kbps)
• 9,000 x 0.05 x (439,000 x 8) / 1,800 = 857.42
• Total throughput
• 131.25 + 436.72 + 857.42 = 1,425.39 Kbps
8 Menascé & Almeida. All Rights Reserved.
31
Quality of Service
• As Web sites become a fundamental component of
businesses, quality of service will be one of the top
management concerns.
• The quality of the services provided by a Web
environment is indicated by its service levels,
namely:
• response time
• availability
• reliability
• cost
8 Menascé & Almeida. All Rights Reserved.
32
Quality of Service
• To a user, the source of problems is immaterial
• users want performance, security, 24H x 7D uptime
• The problem of quality of service on the Web is
exacerbated by the unpredictable nature of
interaction of users with Web services. It is usual to
see the load of a Web site being multiplied by 8 on
the occurrence of a special event.
• How does management establish the service levels
of a Web site?
8 Menascé & Almeida. All Rights Reserved.
33
Quality of Service
• Typical questions to help to establish the service level
of a Web service:
– Is the objective of the Web site to provide
information to external customers?
– Do your mission-critical business operations
depend on the World Wide Web?
– Do you have high-end business needs for which
24 hours-a-day, 7 days-a-week uptime and high
performance are critical, or can you live with the
possibility of Web downtime?
8 Menascé & Almeida. All Rights Reserved.
34
QoS metrics examples
•
•
•
•
Dial-up call failure rates
Time it takes to log in
Web site throughput
Web site time-out rate
• Capacity planning for peak periods is
fundamental, due to high variability of
workload
8 Menascé & Almeida. All Rights Reserved.
35
Web infrastructure
INTERNET - TCP/IP INFRASTRUCTURE
Public
Web Site
HTTP Server
Firewall +Cache/Proxy
O.S. - TCP/IP
Hardware
Private
Web Site
HTTP Server
O.S. - TCP/IP
Hardware
Intranet
TCP/IP
Desktop Computer
Desktop Computer
Browser
Browser
O .S.
Network
Desktop Computer
Hardware
Browser
O .S.
Network
Hardware
8 Menascé & Almeida. All Rights Reserved.
O .S.
Network
Hardware
Mirror
Web Site
HTTP Server
O.S. - TCP/IP
Hardware
Web Proxy Architecture
Clients
Proxy
Server
8 Menascé & Almeida. All Rights Reserved.
Servers
37
Web Proxy Architecture
• Originally introduced for users behind a Firewall.
• A proxy acts as an agent, representing the server to
the client and the client to the server.
• A proxy accepts request from clients and forwards
them to Web servers.
• Once a proxy receives responses from remote
servers, it passes them to clients.
• Proxies can be configured to cache relayed
responses, becoming then a caching proxy.
8 Menascé & Almeida. All Rights Reserved.
38
Web Caching Proxies (I)
• Caches are always used hierarchically:
– OS / disk controller; browser (local)
– caching proxy (network)
• Caching improves speed and network costs,
bringing data closer to users + reducing
server load
• Metrics for cache effectiveness:
– (document) hit ratio
– byte hit ratio
– data transferred
8 Menascé & Almeida. All Rights Reserved.
39
Web Caching Proxies (II)
• Caching adds its own set of problems:
– how to be sure that document is up-to-date?
(cache consistency, or cache coherence):
• Expires MIME header (optional)
• HEAD HTTP request
• GET if-modified-since - 304 status mesg. (Not Modified)
– what documents are worth caching?
– For how long?
8 Menascé & Almeida. All Rights Reserved.
40
Web Caching Proxy: an example
• A large company decided to install a caching proxy
server on the corporate intranet. After 6 months of
use, management wanted to assess the caching
effectiveness. So, we need performance metrics to
provide quantitative answer for management.
• Cache A: we have a cache that only holds small
documents, with average size equal to 4,800 bytes.
The observed hit ratio was 60%.
• Cache B: the cache management algorithm was
specified to hold medium documents, with average
size of 32,500 bytes. The hit ratio was 20%
8 Menascé & Almeida. All Rights Reserved.
41
Web Caching Proxy: an example
• The proxy was monitored during 1 hour and 28,800
requests were handled in that interval.
• Let us compare the efficiency of the two cache
strategies by the amount of saved bandwidth
• SavedBandwidth = (No-of-Req.Hit-Ratio Size)/Int.
• SavedBandwidth-A = (28,800 0.64,800 8)/3,600
= 180 Kbps
• SavedBandwidth-B = (28,800 0.232,500 8)/3,600
= 406.3 Kbps
8 Menascé & Almeida. All Rights Reserved.
42
Web Mirrors
• Implies replicating site contents at other
servers
• Requires regular, usually automatic
redistribution of contents updates
• Best combined with DDNS (dynamic DNS) to
redirect browsers
• Purpose:
– increase availability, minimize geographical
distances
– balance server load (DDNS)
8 Menascé & Almeida. All Rights Reserved.
43
Web Server Architecture
• Need to service more requests at a time
– fork a copy per client
– multi-thread server program
– pool of running services
• Need to split traffic among machines
– DDNS (round-robin; other strategies)
– single-IP-image
– dynamic HTTP redirection
8 Menascé & Almeida. All Rights Reserved.
44
Workload of Web servers (I)
• Has unique characteristics:
– dynamic WEB pages
• client-side programs (applets)
– JAVA, ActiveX, Jscript, Vbscript, Dynamic HTML
• server-side programs (servlets)
– can severely impact on server performance
– JAVA, ActiveX, PERL, C/C++, Shell, ASP, PHP
– CGI - Common Gateway Interface, universal
interface for non-embedded applications
– ISAPI (W/NT only) uses threads to reduce overhead
8 Menascé & Almeida. All Rights Reserved.
45
Workload of Web servers (II)
• Has unique characteristics (2):
– extreme variability
• notion of class of requests (clustering) - e.g.
– small pages (less than 5 kB)
– medium documents (5 to 50 KB)
– large files (more than 50 KB)
• Zipf’s law:
f ~ 1/r
• f = frequency of use; r=rank of popularity
8 Menascé & Almeida. All Rights Reserved.
46
Workload: dynamic Web pages
• The Web site of a virtual bookstore receives an
average of 20 visitors per second. One out of 10
visitors places an order for books. Each order
transaction generates a CGI script, which is executed
on the Web server. The Webmaster wants to know
what is the CPU load generated by the CGI script.
• Consider that the average CPU service demand of a
CGI script is: Dcpu = 120 msec./CGI
• Using the Service Demand Law:
• Ucpu = Xcgi  Dcpu
8 Menascé & Almeida. All Rights Reserved.
47
Workload: dynamic Web pages
• Xcgi = cgi = VisitRate  PercentageOfOrders
= 20  (1/10) = 2 CGI/sec
• Ucpu = 2  0.12 = 0.24 = 24%
• What would be the impact of replacing the CGI
applications by servlets? Let us assume that Java
servlet transactions are 30% less resource-intensive
than CGI applications.
s
cpu
D
= D  0.7 = 120  0.7 = 84 msec.
cgi
cpu
• The CPU utilization due to servlets would be
Ucpu = 2  0.084 = 0.168 = 16.8%
8 Menascé & Almeida. All Rights Reserved.
48
Novel Features in the WWW
• The Web exhibits extreme variability in workload
characteristics:
– Web document sizes vary in the range of 102 to 106 bytes
– The distribution of file sizes in the Web exhibits heavy tails. In
practical terms, heavy-tailed distributions indicate that very
large values are possible with non-negligible probability.
– This reduces statistical meaning of measures, which can be
improved by using request classes.
• Web traffic exhibits a bursty behavior
– Traffic is bursty in several time scales (factors of 8 to 10).
– It is difficult to size server capacity and bandwidth to support
demand created by load spikes.
8 Menascé & Almeida. All Rights Reserved.
49
Novel Features in the WWW
• The manager of the Web site of a large publishing
company is planning the capacity of the network
connection.
• 1 million HTTP operations per day
• average document requested was 10 KB
• The required bandwidth (Kbps) is:
HTTP op/sec  average size of documents (KB)
11.6 HTTP ops/sec  10 KB/HTTP op = 928 Kbps
• Assume that protocol overhead is 20%
8 Menascé & Almeida. All Rights Reserved.
50
Novel Features in the WWW
• The actual throughput required is
928  1.20 = 1.114 Mbps
which can be provided by a T1 connection.
• Assume that management decided to plan for peak
load. The hourly peak traffic ratio observed was 5 for
some big news event. Then the required bandwidth
is:
1.114  5 = 5.57 Mbps
which requires four T1 connections.
8 Menascé & Almeida. All Rights Reserved.
51
Capacity Planning of Web Servers
• It can be used to avoid some of the obvious and most
common pitfalls: site congestion and lack of
bandwidth.
• Typical capacity planning questions:
• Is the corporate network able to sustain the
intranet traffic?
• Will Web server performance continue to be
acceptable when twice as many people visit the
site?
• Are servers and network capacity adequate to
handle load spikes?
8 Menascé & Almeida. All Rights Reserved.
52
Part III: Summary
•Web server problems
•Combination of HTTP and TCP/IP
•Basic examples using operational analysis
•Bottlenecks
•Perception of performance and metrics
•Quality of Service
•Web caching proxy
8 Menascé & Almeida. All Rights Reserved.
53