Upload
akashag11111
View
217
Download
1
Tags:
Embed Size (px)
Citation preview
The Internet grows, and grows…
0
1020
3040
5060
7080
90
Jan-97 Jul-97 Jan-98 Jul-98 Jan-99 Jul-99 Jan-00 Jul-00
Number of hosts (millions) in the Internet (Netsizer by Telcordia)
Network versus servers: who is trailing whom?
• Network vs. server performance, today
• State of the network,• The special case of the DNS,• The 10 Gigabit challenge,• Going end-to-end?
Network vs. server performance, today
get address
connect
DNS
accept
prepare
transmit
HTTP GET
receive
close
server
network
delay
Get address only once (cache).Connect only once (HTTP 1.1), Pipeline.
Measuring the components of the web
delayDNS
Web performance estimation, 1998-2000
• For a given client/server pair:– Measure the address resolution delay,
connection delay, delay to first packet, delay to transmission of page
• Repeat every day:– For a fixed set of 100 “large” servers,– For a set of 100 “random” servers,
provided from randomized requests to Altavista, Lycos, Google, etc.
Observed bandwidth seems to grow over time
0
20000
40000
60000
80000
100000
120000
140000
Sep-98 Feb-99 Aug-99 Jan-00
median
worst 5%
14 per.Mov. Avg.(median)
The servers seem to contribute 40% of the
delay
0%
10%
20%
30%
40%
50%
12/1/99 1/1/00 2/1/00
Large servers Random servers
Feedback Loop: Why Capacity Doubles Every 8
Months
Providers increase the network capacity
More users, more demand,
faster modems
More contents, more servers,
faster feeds
Network vs. server: what is the state of the
network?
First network characteristic:
Packet loss ratesEvolution of SYN repeat rates
0%
5%
10%
15%
20%
25%
9/4/1998 2/14/1999 7/27/1999 1/6/2000
C > 2s
14 per. Mov. Avg. (C >2s)
Second characteristic:Round trip time
0
50
100
150
200
250
300
Jul-98 Nov-98 Feb-99 May-99 Aug-99 Dec-99 Mar-00
Taking a closer look at the round trip time distribution
Density and CFD of "short" connection delays (ms)
0.000001
0.00001
0.0001
0.001
0.01
0.1
1
10 100 1000 10000
1-CFD(x)
density(x)
Likely evolution of the network
• Accommodate growing demand:– Cable, DSL, other broadband– Mobile phones, appliances
• Likely evolution:– Average bandwidth grows -> 100 kbps– Median delays -> distance/speed of light– Delay deviation diminishes, as predicted
by Paul Kenny’s work.
Network vs. server:The special case of the
DNS
DNS: an amplifier of the Internet Performance
Local S
erv
er Root, .Com
Example.Com
Many network traversals, many servers.
Resolution Delay,Random DNS Names
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0 2 4 6 8 10 12
DNS delays don’t seem to improve at Internet
Speed…
0%
10%
20%
30%
40%
50%
60%
9/4/1998 2/14/1999 7/27/1999 1/6/2000
D > 2s
C > 2s
14 per. Mov. Avg. (D >2s)
14 per. Mov. Avg. (C >2s)
Fraction of DNS resolution delays over 2 seconds
A clear case of “server vs. network”
• A.ROOT-SERVERS.NET loss: 33.9%• B.ROOT-SERVERS.NET loss: 3.4%• C.ROOT-SERVERS.NET loss: 21.1%• D.ROOT-SERVERS.NET loss: 2.6%• E.ROOT-SERVERS.NET loss: 12.3%• F.GTLD-SERVERS.NET loss: 4.0%• F.ROOT-SERVERS.NET loss: 12.3%• G.ROOT-SERVERS.NET loss: 62.1%• H.ROOT-SERVERS.NET loss: 20.9%• I.ROOT-SERVERS.NETloss: 7.1%• J.GTLD-SERVERS.NET loss: 8.3%• K.GTLD-SERVERS.NET loss: 2.2%
Limits of these data:
One single test,
One single day!
How will the DNS situation evolve?
• No clear virtuous cycle:– Registration of new names is a for profit
venture (NSI, and wannabes)– Maintenance of root is a non profit
service with political constraints (ICANN)• Solutions from servers, software:
– Less reliance on root?– More reliance on caches (load
balancing?)– Less reliance on the DNS!
Network vs. server:The 10 Gigabit challenge
Servers are getting better but is this enough?
• Internet2 Land Speed Record award:– "We hope this competition gets people
thinking about enabling really revolutionary Internet applications," - Jim Gray
– "To realize Internet's full potential, end-to-end network performance needs to take a huge leap forward," - Gordon Bell
• The records: – 751.362 Mbps over 5,626 km (single TCP)– 957.369 Mbps over 5,626 km (2 TCP)
The 10 Gigabit challenge, or the limits of Moore’s
law
02468
1012141618
2000 2001 2002 2003
Backbone (9months)
Fiber (10months)
Moore's law (18months)
Meeting the challenge:part 1, ever better servers• Get stronger:
– Faster CPU, better memory architecture– Faster I/O
• Get smarter:– HTTP 1.1, TCP, IPv6…– Better protocol design(s)– Better software (doing that)
• Get help:– Offload packet processing (done that)– Offload encryption (done for IPSEC)– Offload key management (uh?)
Meeting the challenge:part 2, larger clusters
• Servers x2 every 18 month. • Can we manage a cluster of 100,000
servers?– Load balancing, Software upgrade, Failover
• Can we develop distributed applications?
Meeting the challenge:part 3, distribution
networks
• Push the “content” near the “consumers”
• Is this needed in a 10 Gigabit world?• Does it work for more than GIF files?
Evolution of server, server farms and distribution
• Handling variability– Heavy tail distribution of demand means
flash crowds, etc.– Could incite to a “market of servers”, or to
“adaptive servers.”
• Pushing up hill– Need a serious business case for investment
in new servers– Constantly testing the limits of the
technology
Network vs. server:Going end-to-end?
End to end distribution:NAPSTER, Gnutella
Rendezvous, directory
End to end communication: real time
audio, video, gamesSIP proxy
SIP proxy
Meeting the challenge:distribute the load!
• The return of the end-to-end argument– Each new client brings its own
resource– The only known way of scaling
• Only place in the servers the functions that cannot possibly placed in the client!
Network vs. server:Thank You!