Performance and Robustness Testing

1
Performance and Robustness Testing of Wireless
Web Servers Guangwei Bai Kehinde Oladosu Carey
Williamson
2
1. Introduction and Motivation

• Observation the same wireless technology
• that allows a Web client to be mobile also
• allows Web servers to be mobile
• Idea portable, short-lived, ad hoc networks
• Possible applications
• classroom area networks, seminars
• press conferences, media events
• sporting events, gaming, exhibitions
• conferences and trade shows
• disaster recovery sites, field work, etc.

3
Background Portable Networks

• Assumptions the characteristics of a
• portable short-lived network are
• set it up when needed tear down after
• only needed for minutes or hours
• when may not be known a priori
• where may not be known a priori
• no existing infrastructure of any kind
• general Internet access not available
• general Internet access not required
• pre-defined content target audience
• 1-100 users mobile limited bw needed

4
2. Objectives

• to assess feasibility of portable networks
• to benchmark the performance capabilities
• and limitations of an Apache Web server in
• a wireless ad hoc network
• to identify the performance bottlenecks
• to understand impacts of different factors
• number of clients
• Web object size
• persistent connections
• transmit power (energy consumption)
• wireless channel conditions

5
3. Experimental Setup

• Compaq Notebooks (1.2GHz Pentium III, 128MB RAM,
  
• 512 KB L2 cache, Cisco Aironet 350 network
  cards)
• RedHat Linux 7.3, httperf, Apache 1.3.23,
  SnifferPro 4.6
• Network 11 Mbps IEEE 802.11b wireless LAN, ad
  hoc mode

6
Experimental Setup (Contd)

• IEEE 802.11b a standard for wireless LANs
• Carrier Sense Multiple Access with Collision
  Avoidance
• (CSMA/CA), up to 11 Mbps data rate at physical
  layer
• ad hoc mode
• frames are addressed directly from sender to
  receiver
• httperf
• Web benchmarking software tool developed at HP
  Labs
• Web server Apache (version 1.3.23)
• Process-based, flexible, powerful,
  HTTP/1.1-compliant
• SnifferPro 4.6
• real-time capture, recording all wireless channel
  activity,
• enabling protocol analysis at MAC, IP, TCP and
  HTTP layers

7
 
 
4. Experimental Design

• Impacts of different factors on wireless Web
  server
• performance (one-factor-at-a-time)

Experimental Factors and Levels

• Performance metrics
• HTTP transaction rate, throughput, response
  time, error rate
• at Application Layer,
• TCP connection duration at Network Layer
• Transmit queue behaviour at Link Layer,

 
 
8
5. Measurement Results and Analyses -
Expt 1 Request Rate - Expt 2 Transfer
Size - Expt 3 Number of Clients -
Expt 4 Persistent Connections - Expt 5
Transmit Power - Expt 6 Wireless Channel
9
Experiment 1 Request Rate

• Purpose to determine the range of feasible and
  sustainable
• loads for the wireless Web
  server
• Design
• Number of Clients 1
• HTTP transaction rate 10, 20, , 160 req/sec
• HTTP transfer size 1 KB (fixed)
• Persistent connections no
• Transmit power 100 mW
• Client-server distance 1 meter (on same desk)

10
Wireless Web Performance at Application Layer

• Main observation
• As the offered load increases
• linear increase ? instability ? lower
  plateau
• Peak throughput lt 1 Mbps for 1 KB transfers

11
Transmit Queue Behaviour for Experiment 1

• Main observation Wireless LAN is the bottleneck
• Packet drops occur from link-layer queue (client
  side)
• Even before they get on the wireless LAN!!!

• Reason
• No flow control / backpressure mechanism
• Note default queue size is 100 in the Linux
  kernel

12
Wireless Web Performance at Application Layer
(Contd)

• Main observation
• the response time is about 9 ms at low load,
  increase
• significantly to over 2 sec at high load (gt85
  req/sec)
• failures occur frequently under overload

13
Measurement at Network Layer
14
Experiment 2 Transfer Size

• Purpose to study impact of HTTP response size
• Design
• Number of Clients 1
• HTTP transaction rate 10 req/sec (fixed)
• HTTP transfer size (KB) 1, 2, 4, 8,
• Persistent connections no
• Transmit power 100 mW
• Client-server distance 1 meter (on same desk)

15
Measurement at Network Layer
General observation as HTTP transfer
size increases, mean TCP connection duration
increases, as does the variance of distribution.
16
Measurement at Network Layer
17
Experiment 3 Number of Clients

• Purpose to study impact of high load generated
  by
• multiple clients
• Design
• Number of Clients 2, 3, 4
• HTTP transaction rate 10, 20, , 160 req/sec
• HTTP transfer size 1 KB (fixed)
• Persistent connections no
• Transmit power 100 mW
• Client-server distance 1 meter (on same desk)

18
Wireless Web Performance at Application Layer (4
Clients)
19
Wireless Web Performance at Application Layer (4
Clients)

• Main observation
• 4 clients share network and server resources
  equally
• 30 higher aggregate throughput (110 conns/sec)
• bottleneck is now at server network card
  (drops!!)

20
Wireless Web Performance at Application Layer (2
or 3 Clients)
21
Wireless Web Performance at Application Layer (2
or 3 Clients)
Main observation unfairness problem at high
loads one client obtained a higher proportion
of the throughput at expense of another (dont
know why?)
22
Experiment 4 Persistent Connections

• Persistent Connections
• Multiple HTTP transactions can be sent on the
• same TCP connection.
• amortize overhead of TCP connection processing
• reduce memory consumption for TCP state
• Purpose of this experiment to study impact of
• persistent connection on wireless Web
  performance
• Design
• Number of Clients 1 and 2
• HTTP transaction rate 10 req/sec (fixed)
• HTTP transfer size 1 KB (fixed)
• Persistent connections yes
• Transmit power 100 mW
• Client-server distance 1 meter (on same desk)

23
Achieved Throughput for Experiment with
Persistent Connections

• Main observation
• Peak throughput 3.22 Mbps, 3.5x improvement
• over non-persistent connections (0.9 Mbps),
• two clients share the server and network
  resources
• equally

24
Experiment 5 Transmit Power

• Energy consumption- an important issue for mobile
• Clients and Server.
• Purpose to see what transmit power is required
  for
• acceptable performance in
  classroom setting
• Design
• Number of Clients 1
• HTTP transaction rate 10 req/sec (fixed)
• HTTP transfer size 1 KB (fixed)
• Persistent connections no
• Transmit power 1, 5, 20, 100 mW
• Client-server distance 10 meter (same floor)

25
Measurement at Network Layer

• General observation
• If transmit powerlt10 mW
• MAC-layer retransmits
• rightward skew
• unacceptable perf.
• If transmit power?20 mW
• acceptable performance

26
Experiment 6 Wireless Channel Characteristics

• Wireless Internet is characterized by limited
• bandwidth, high error rates, and interference.
• Purpose to study the impact of the wireless
  channel
• characteristics on wireless
  Web performance
• Design
• Number of Clients 1
• HTTP transaction rate 10 req/sec (fixed)
• HTTP transfer size 1 KB (fixed)
• Persistent connection no
• Transmit power 100 mW
• Client-server distance 1m, 10m, 100m

27
Measurement at Network Layer (100m scenario)
Low load 10 req/sec Significant skew to the tail
of the distribution, Some periodicity (why?)
Medium load 50 req/sec Significant skew to
the tail of the distribution
28
6. Summary and Conclusions

• What we did wireless Web server, portable nw
• Application-layer measurements (httperf)
• Network-layer measurements (Wireless Sniffer)
• Our results show
• Server capability 100 conn/sec for
  non-persistent
• HTTP with throughputs up to 4 Mbps
  (adequate?)
• Bottleneck at wireless network interface
• Some network thrashing for large HTTP
  transfers
• when the network utilization is high (aborts,
  resets)
• Effect of wireless channel on performance at
• TCP and HTTP-level (MAC-layer retransmits)
• Power consumption issue for mobile client and
  server

29
7. Future Work

• Explaining the anomalies (fairness, periodicity)
• Better system instrumentation (Linux)
• More realistic Web workloads
• Larger WLAN testing (classroom scenario)
• Repeat experiments with IEEE 802.11a (55 Mbps)
• Kennys M.Sc. Thesis...
• Another paper?