Transcript Capacity Planning for the Internet

Capacity Planning for
Internet Service
Networks
Geoff Huston
NTW Track4
Issues
TCP/IP Protocol Behavior
Issues
Usage Profile
Capacity Guidelines
Growth Levels
Planning Issues
Understand the domain of operation
technical issues
market issues
competitive issues
regulatory issues
TCP/IP Protocol Issues
TCP/IP is NOT a flow damped protocol
end to end flow management
sliding window protocol
adaptive flow rate designed to probe and use
max available end to end bandwidth
only limited by end system buffering size
bandwidth x delay
system buffers are getting larger as OS
vendors come to understand the problem
TCP/IP Protocol Issues
TCP/IP Data Flow Rate Adaptation
Steady State Available Bandwidth
Data Flow Rate
Rate overflow loading
into network queues
Time
TCP/IP Protocol Issues
No network-based flow control mechanism
Network-based packet loss signals end systems
to collapse transmission window size
Varying window size allows adaptive flow
metrics to adapt to changing maximum available
capacity
Sustained insufficient capacity leads to
congestion induced collapse of data throughput
TCP/IP Protocol Issues
Many simultaneous TCP sessions interact
with non-predictive non-uniform load
(ftp://thumper.bellcore.com/pub/dvw/sigcom93.ps.Z)
Peaks start to synchronize with each other
Buffering evens out individual flows, but
buffers themselves behave with
fluctuating load
Buffering adds latency
TCP/IP Protocol Issues
66%
Congestive Collapse The slide to misery
and packet loss
33%
Data Throughput
TCP/IP efficiency under congestion load
33%
Traffic Level
66%
TCP/IP Protocol Issues
TCP vs UDP
UDP-based applications
Internet Phone, Video, Workgroup
UDP Issues
no flow control mechanism
sustained use forces precedence over TCP
flows
increasing use of flow bandwidth negotiated
protocols for these applications (RSVP)
TCP/IP Protocol Issues
Damping network capacity is not a
demand management tool
Network capacity must be available to
meet peak demand levels without
congestion loss
Usage Profile
Two major Internet use profiles:
Business use profile
peak at 1500 - 1600
plateau 1000 - 1730
Residential dial profile
peak at 2030 - 2330
plateau 1900 - 2400
Usage Profile
Distance profiles
12% Local
18% Domestic Trunk
70% International
Traffic mix due to:
Distance invisible applications without user
control
Distance independent user tariff
Capacity Guidelines
Link Utilisation
Average weekly traffic level set to 50% of
available bandwidth.
Core network capacity should be
dimensioned according to aggregate
access bandwidth
Link Usage Profile - optimal
peak loading less than 10% time
greater than 50% loading for 50%
time
traffic bursting visible
Link Usage Profile - overloaded
90% peak loading for 45% time
60% peak loading for 60% time
no burst profile at peak loads
imbalanced traffic (import based)
Link Usage Profile - saturated
visible plateau traffic load signature
small load increases cause widening
plateau
Overall Growth Levels
Two growth pressures:
serviced population
the changing Internet service model
more network-capable applications
using more bandwidth
Overall Growth Levels
For a constant service model the growth
curve will exhibit demand saturation
500
400
300
200
100
0
1
11
21
31
41
51
61
Overall Growth Levels
 For a changing service model the saturation point will
move
 More intense network use by increasingly sophisticated
applications
1200
1000
Technology shift
800
600
400
200
0
1
11
21
31
41
51
61
How to plan
Generate a market demand model
forecast the number of services in operation
existing services
growth rate
market capture level (competitive position)
forecast the average demand per service
dial access, leased line
web, ftp, usenet
caching trends
new Internet services
How to plan
demand models are typically very uncertain
indicators
high level of uncertainty of externalities
highly dynamic competitive position
poorly understood (and changing) service
demand model
How to plan
Forward extrapolation
assume existing traffic follows a general
growth model
forward extrapolat the growth model
Good for short term planning (12 months)
Cannot factor
latent demand
market price sensivity
4/27/97
4/17/97
4/7/97
3/28/97
3/18/97
3/8/97
2/26/97
2/16/97
2/6/97
1/27/97
1/17/97
1/7/97
12/28/96
12/18/96
12/8/96
11/28/96
11/18/96
11/8/96
10/29/96
10/19/96
10/9/96
9/29/96
9/19/96
9/9/96
8/30/96
8/20/96
8/10/96
7/31/96
7/21/96
7/11/96
7/1/96
Trend forecasting
 historical usage vs capacity data
120000
100000
80000
60000
Total Capacity
Daily Traffic IN
Daily Traffic OUT
40000
20000
0
23/8/98
27/7/98
30/6/98
3/6/98
7/5/98
10/4/98
14/3/98
15/2/98
19/1/98
23/12/97
26/11/97
30/10/97
3/10/97
6/9/97
10/8/97
14/7/97
17/6/97
21/5/97
24/4/97
28/3/97
1/3/97
2/2/97
6/1/97
10/12/96
13/11/96
17/10/96
20/9/96
24/8/96
28/7/96
1/7/96
Growth Trends
600000
500000
400000
in
out
Trend - high
Trend - mid
300000
Trend - Low
97/98 bandwidth
200000
100000
0
Planning
undertake demand and trend forecast
models
constantly review the model against
generated usage data
recognise that the larger the capacity you
need the longer the lead time to purchase it
recognise that the bigger the purchase the
greater the requirement for capital