Evaluating the Impact of DNS and HTTP Session Characteristics on Consumer ISP Web Traffic Jason But, Urs Keller and Grenville Armitage Centre for Advanced Internet Architectures Swinburne University of Technology
[email protected] November 2005
Outline
• Motivation • Experimental Environment • What makes up a HTTP Session • Results • •
Gathered DNS Lookup Statistics Gathered HTTP Transaction Statistics
• Caching Content • •
Cacheability of Content Potential to Improve the Online User Experience
TenCon2005
http://caia.swin.edu.au
[email protected]
November 2005
Page 1
Motivation
• Web caches considered good • Evaluate their effectiveness in the consumer ISP context • Aim to meaure the entire user web experience • •
•
Correlate HTTP transactions with corresponding DNS queries Measure the impact of client-server RTT with respect to HTTP transaction time Estimate potential improvements to user web access times using different caching models
TenCon2005
http://caia.swin.edu.au
[email protected]
November 2005
Page 2
Experimental Environment
DSL Modem i n b r i dgi ng m ode Net sn i f f Box Hom e Net wor k
Int er net
• NetSniff1 Box running FreeBSD 5.3 • • • • •
1
Manages ISP Connection Provides DHCP and DNS forwarding services within the home LAN Provides a NAT service to any computer on the home LAN Restarts data gathering upon failure Logs rolled over daily and transfered to CAIA at 3AM
Multi protocol layer network traffic analysis tool developed at CAIA - http://caia.swin.edu.au/ice/tools/netsniff TenCon2005
http://caia.swin.edu.au
[email protected]
November 2005
Page 3
What Contributes to a HTTP Session
DNS Requ est
∆t DNS l ook u p
DNS Response
• DNS Query ∆t thi nk − ti m e
SYN
•
If result is not locally cached
∆t ha ndsha k e 1
SYN/ ACK
• HTTP Transaction ACK HT T P Requ est
•
∆t ha ndsha k e 2
HT T P Response
•
TCP Handshake Content Transfer Time
∆t conte nt transfe r Client
Server
TenCon2005
http://caia.swin.edu.au
[email protected]
November 2005
Page 4
DNS Lookup Statistics
DNS Lookup Times for HTTP Transactions having a lookup 407
276
Cum. Distrubtion 0.4 0.6
0.8 Cum. Distrubtion 0.4 0.6
95% quantile mean
0.2
median 95% quantile mean
0.2 0.0
57
1.0
137
0.8
1.0
40
DNS Lookup Times for HTTP Transactions including browser cached entries
0
200
400 Observed Request Time [ms]
600
0
800
100
200 300 400 Observed Request Time [ms]
500
• About 73,500 DNS Requests
• Total 179,000 HTTP Transactions
• Requests within 40ms handled by ISP DNS Server
• Other transactions incur no DNS lookups
• Responses > 250ms reflect access to an overseas DNS
TenCon2005
•
•
DNS resolution cached at client Equivalent to a DNS resolution of 0ms
http://caia.swin.edu.au
[email protected]
November 2005
Page 5
Hop Counts to HTTP Servers
Path Hop Count Distribution for HTTP Transactions 1.0
18
24
0.2
Cum. Distribution 0.4 0.6 0.8
median 95% quantile mean
0
5
10
15 20 Observed Hopcount
25
30
35
• Results indicate the absence of a Web Cache at the ISP • Mean/median hop count to a web server is 18
TenCon2005
http://caia.swin.edu.au
[email protected]
November 2005
Page 6
RTT Distributions
630
RTT Jitter
time [ms] 300 400
500
Cumulative Distribution 0.2 0.4 0.6 0.8
600
1.0
320
700
RTT Distribution for HTTP Transactions
Mean RTT and Jitter vs. Path Hop Count for HTTP Transactions
200
400 600 Observed Round Trip Timea [ms]
800
1000
0
0
100
200
median 95% quantile mean
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35
path hop count
• Resonably linear spread of HTTP RTT distributions •
Interrupted at 200ms (overseas servers)
• RTT and Jitter both related to number of hops in path
TenCon2005
http://caia.swin.edu.au
[email protected]
November 2005
Page 7
Would a Web Cache Help
• Decrease the Hop Count to the server • Decrease the 2 x RT T factor that makes up the overall content download time •
Which will remain constant even as Internet access rates increase
• Minimise DNS query time • Is enough content cacheable to justify? • How much of an impact would it have?
TenCon2005
http://caia.swin.edu.au
[email protected]
November 2005
Page 8
Cacheability of Content
Cacheability of content types as a proportion of total HTTP Transactions
Cacheability of content types as a proportion of total downloaded bytes text/plain [0%,27.89%]
image/gif [3.26%,39.25%]
application/x−tar [0%,16.01%]
text/html [5.06%,14.51%]
image/jpeg [0.03%,13.1%]
image/jpeg [0.07%,18.34%]
text/html [0.53%,7.84%]
application/x−javascript [1.17%,3.43%]
image/gif [0.07%,7.41%]
text/css [0.04%,1.81%] image/png [0.06%,1.31%] previously cached [0%,8.99%] other [0.63%,2.07%] (13 different MIME types witnessed)
application/x−gzip [0%,5.35%] video/quicktime [0%,4.39%]
Non−Cacheable Cacheable
Non−Cacheable Cacheable
video/mpeg [0%,3.26%] application/pdf [0%,2.24%] other [0.13%,11.75%] (10 different MIME types witnessed)
• Majority of downloaded content is cacheable • •
~ 90% of individual HTTP transactions > 99% of downloaded bytes
• Whether content is cached depends on the hit-rate of the cache • Potential for large savings in access times TenCon2005
http://caia.swin.edu.au
[email protected]
November 2005
Page 9
Potential Benefits
Characteristics of HTTP Transactions
SYN SYN/ ACK
800 time [ms] 400 600
∆t DNS l ook u p
200
∆t thi nk − ti m e
0
DNS Response
Cacheable
∆t ha ndsha k e 1
ACK HT T P Requ est HT T P Response
∆t ha ndsha k e 2 ∆t conte nt transfe r
Client
Transfer Time First Data from Server First SYN, SYN/ACK Think time DNS lookup
1000
DNS Requ est
Server
TenCon2005
Non Cacheable
Cacheable passive Cache 30% hit−rate
Cacheable prefetch Cache 40% hit−rate
• Transfer time calculated as download at 512kbps • For a cache hit • DNS lookup will be 0ms • HTTP RTT is equal to 20ms • More potential for savings: • Up to 88% of content is cacheable
http://caia.swin.edu.au
[email protected]
November 2005
Page 10
Conclusions
• ISP’s we used did not cache web content locally • RTT to the web server forms a significant proportion of the HTTP Transaction time • Web Caches would reduce this RTT •
Potential for 24-32% savings in average HTTP transaction time
• Reduce core traffic for other Internet applications • Potential to become more important as access speeds increase • •
Data transfer time drops RTT to server becomes a greater part of the HTTP transaction time
TenCon2005
http://caia.swin.edu.au
[email protected]
November 2005
Page 11
