Documente Academic
Documente Profesional
Documente Cultură
Computer Networks
CS 552 Professor: Badri Nath
http://www.cs.rutgers.edu/~badri
badri@cs.rutgers.edu
Badri Nath Office Hours: Wednesday 1:30-3:30 PM; CORE 320
TA: Long T Le
Rutgers University
longtle@cs.rutgers.edu
badri@cs.rutgers.edu Office Hours: Tuesday 10:00-11:00 AM; Hill 202
SPRING 2017 Course info
http://www.cs.rutgers.edu/~badri/552.html
1 2
https://sakai
Course schedule Will submit reviews or answer to mini-quizzes
Reading list (based on the paper) online
Lecture notes Course announcements
Announcements Written Homeworks
Assignments
Project ideas
Exams
3 4
1
Course Goals
Course Materials
Understand /identify the basic idea Computer Networks: A Systems Approach, 4th Ed. (2007), by
Larry Peterson and Bruce Davie.
What is the problem that the paper tackles?
Computer Networking: A Top-Down Approach Featuring the
What kind of a paper? Internet, 5th Ed. (2010), by James F. Kurose and Keith W. Ross.
Performance, vision, new direction/protocol paper TCP/IP Illustrated, Volume 1: The Protocols by W. Richard
Summarize key idea Stevens.
Unix Network Programming: Networking APIs: Sockets and
+ve aspects of the paper XTI (Volume 1) by W. Richard Stevens.
New, breakthrough, incremental, Advanced Programming in the Unix Environment by W.
-ve aspects of the paper Richard Stevens, Addison-Wesley, 1993.
Readability, Assumptions (valid?), scaling issues (does it I or 2 recommended
scale), implementation (has it been implemented),
measurements (problems?)
7 8
2
Grading
Class Coverage
3
What is a Network? Why Networks?
13 14
What is an internet?
Network of networks
What is the Internet?
A global internet based on the IP protocol
To what does Internet technology refer?
Architecture, services, interfaces, and
protocols
15 16
4
Impact on society
Net neutrality
ISPs should not discriminate
For and against case
20 M msgs/ minute Nations laws and Internet
Regulation across national boundaries
Privacy
2.4 M Queries/ minute Ears and Eyes in your house connected to the net.
Who can ask for that?
Scale Scale Scale Is Alexas (Amazon) recordings available to others?
Content creation , ownership, distribution, online piracy
Cyber Warfare N Korea, Hacking (political objectives)
17 18
19 20
Mary meeker : Internet trends May 2015
5
Internet Growth WWW growth
21 22
Souce: Internet trends May 2013, Mary Meeker: Astounding growth of Google
23 24
6
Facebook growth Social Networking/Messaging
25 26
MAU: Number of unique users in the past 30 days
27 28
7
Now: everyone; grandpa, grandma Mobile Phone usage
29 30
Mobile OS trends
31 32
8
Communication trends Video Content Growth
35 36
9
Reimagining shared economy Whats hot?
Embedded
Mobile &Wireless Internet of things (IoT)
Internet Bodily of Things (IoB)
Internet
Data centers
Edge vs Cloud
http://a16z.com/2016/12/16/the-end-of-cloud-computing/
10
Device Growth: connected Smart cities/Smart Campuses
Software defined urban platforms
If software is eating the world, networking is enabling it
41 42
43 44
11
How to handle different How to locate a node?
networks?
45 46
Corruption? Link
Need error detection and correction Multiplexing
Reliability Routing
Data lost? Addressing/naming (locating peers)
Overload Reliability
Congestion control
Flow control
Security
Fragmentation
Encryption, authentication
Etc.
47 48
12
ISO OSI Layering Architecture Problems
Host A Host B
Application Application Protocol Application
Seven layers not widely accepted
Layer Layer
Standardized before implemented
Presentation Presentation Protocol Presentation
Layer Layer
Top three layers fuzzy
Session Session Protocol Session
Layer Layer Internet or TCP/IP layering widespread
Transport Transport Protocol Transport
Layer Layer
51 David Clark, The design philosophy of the DARPA internet protocols, 1998 52
13
Implications (cathedral vs # 2 Multiplexed utilization
bazaar)
53 54
55 56
14
#4 distributed management
Facebook DC
https://code.facebook.com/posts/360346274145943/introducing-data-center-
fabric-the-next-generation-facebook-data-center-network
Major elements:
CDN Content distribution networks
IXPs Internet Exchange points
Datacenters
59 60
4x40 G RSW
15
Internet Design Principles End-to-end argument
61 62
63 64
16
e2e argument e2e tradeoffs
65 66
Energy/power
Always-on system consumes a lot of power
67 68
The cost of a cloud: Research Problems in data center networks by Albert Greenberg et.al, CCR
17
Data center traffic (Of Mice and DC network evolution
Elephants): Latency, BW, $ as metric
>10MB
<100KB
71 72
18
Energy-proportional metric
Energy Proportional Computing
Figure 1. Average CPU utilization of more than 5,000 servers during a six-month period. Servers
73 are rarely completely idle and seldom operate near their maximum utilization, instead operating 74
most of the time at between 10 and 50 percent of their maximum
Throughput proportional
Energy Proportional Computing
fabric
util power EE
0.15 0.6 .25
0.4 0.7 .57 Doing nothing well?
Still power
19
Energy/ Power metric
Energy Proportional Computing
If Cost/Green Conscious
Can we do this for Make network elements less power hungry
Networking [1]Gupta & Singh Greening of the Internet SIGCOMM
Infrastructure? Doing nothing Design for 2002
VERY well wide dynamic [2] Energy Proportionality of an enterprise network, Priya
power range and Mahadevan, et.al, Green Networking August 2010
active low power
modes Take advantage of lower power rates elsewhere,
time-of-day
Energy Efficiency = [3] A. Qureshi, R. Weber, H. Balakrishnan, J. Guttag, B.
Utilization/Power
Maggs, "Cutting the Electric Bill for Internet-Scale
Systems" SIGCOMM 2009
Figure 4. Power usage and energy efficiency in a more energy-proportional server. This
server has a power efficiency of more than 80 percent of its peak value for utilizations of
30 percent and above, with efficiency remaining above 50 percent for utilization levels as
77 78
low as 10 percent.
79 80
P=E/T 1W = 1 J/S 1 Kw-H = 3600000 joules =10 x100W bulbs for 1 Hr
20
More Numbers Enterprise networks
81 82
3 reasons:
Low-Energy Hardware Design:
1. Current energy inefficiencies Use hardware components with low power modes of
a) Wired vs. wireless energy costs operation
b) 6 TW-h ~ 1 nuclear reactor Lower the clock frequency of the components, use
c) Extrapolate to World ~140 nuclear reactors DVS and other methods during low demand
2. Enable greater deployment Energy-Aware APIs give control to software
a) Similar connectivity in India would require 4.75% of total Architecture that allows selective powering off
energy budget
3. Enable longer operation times during events of disaster Energy-Aware Protocol Design:
a) Recent Grid failure in NE US/Canada Node-level algorithms for sleeping
b) Frequent power outages in most of the world Route aggregation and other global techniques to
inform devices when and for how long to sleep
83 84
21
Where to save energy in a device ?
Key Questions
85 86
Uncoordinated sleeping:
Decision to sleep based on individual traffic levels alone Energy cost varies by location
Inform nearest neighbors only
Sleep time limited by protocol hello message timer
Energy cost varies by hour of the dat
May trigger network reconfiguration in case of missed protocol
messages Can we push computation to a geographically
Coordinated sleeping: distant place to save energy?
Decision taken on a network-wide basis
Need algorithm to pre-compute the optimal sleep time, but Can we exploit time difference?
computation costs increase
Hello message frequency can be adjusted, may take longer to detect Peak vs non-peak power rates
changes in network topology
Sleep longer intervals, but forwarding tables may be outdated
Reroute all traffic through one route, shutdown other routes
Introduces delay and packet loss in case of sudden traffic burst
87 88
22
Exploit spatial-temporal price Research
variation
Latency The tail at Scale by Dean and Barroso, CACM, Feb 2013, Vol. 56, NO.2
91 92
23
Use redundancy Hedged requests
S1 S2
r5
Hedged requests: 100 servers
r5 Send hedged request after 10 msec delay
Reduced 99.9 percentile latency from 1800 msec
<r5,S1>
to 74 msec
<r5,S2>
Sends only 2% more requests
Send requests to multiple servers (with id of replicas) Tied requests
Server cancels other requests when the request scheduled
40% reduction at the 99.9 percentile latency
Still both requests may be served (queue empty)
Techniques to reduce overhead: Key takeaway: Predictability from unpredictable parts
Send second after some delay (d> RTT of message in Network) Can we do this for network security?
95 96
24
Computing cost declining: own or rent Cloud services
97 98
On-demand: pay by the hour, no upfront Many cost models for compute resources
payment (0.023 per hour to 0.094 per hour) Storage (Amazon S3)
# CPUs, memory size, disk space etc varies Lease: 1 TB/Month 3$ regular; 1.25$ infrequent;
Glacier (archive) 0.70$
Spot pricing
Buy: 1 TB 60$
Bid for spare capacity, bid price < spot price,
process is preempted
Cost models for Database, management,
storage, network, computation
Reserved Pricing
Data mining/analytics (rapidminer)
1 yr to 3 yr commitment; upfront payment Price proportional to number of tuples used
Large instance $549 plus 0.063 per hour
99 100
25
New Metrics Bid strategies
103
An industry with a mainframe-mentality, reluctant to change
104
Slides from Stanford Site: Nick Mckeown, Martin Casado, Scott Shenker et al,
26
The SDN Approach SDN approach
27
Examples Examples
Switching Routing
Switch MAC MAC Eth VLAN IP IP IP TCP TCP Switch MAC MAC Eth VLAN IP IP IP TCP TCP
Action Action
Port src dst type ID Src Dst Prot sport dport Port src dst type ID Src Dst Prot sport dport
* * 00:1f:.. * * * * * * * port6 * * * * * * 5.6.7.8 * * * port6
Switch MAC MAC Eth VLAN IP IP IP TCP TCP Switch MAC MAC Eth VLAN IP IP IP TCP TCP
Action Action
Port src dst type ID Src Dst Prot sport dport Port src dst type ID Src Dst Prot sport dport
port6,
port3 00:20.. 00:1f.. 0800 vlan1 1.2.3.4 5.6.7.8 4 17264 80 port6 port7,
* * 00:1f.. * vlan1 * * * * *
port9
Firewall
28
SDN issues Research
29