Title: Back to the future with circuit switching
1SECTION 11
- Back to the future with circuit switching
2overview
- Multimedia requirements CSC 461
- Best-effort IP service as too-early binding
- RSVP
- Utility Model Optimal Admission Control problem
3Textbook reading
4 Optimal Admission Control forInternet
multimedia trafficthe best way to design a
controller
- Rob Watson , Mostofa Akbar, Doug Johnson,
Shahadat Khan, Ali Shoja - Eric Manning
- New Media Innovation Centre of B.C. and
- PANDA Lab
- University of Victoria
5The problem
- Tech wreck , telecomm spending collapse
- Internet traffic continues to grow
- Too little of it attracts REVENUE! So,
- Got to find some new apps . . .
- Video on demand Blockbuster,
- videoconferencing since 9/11,
- Interactive video . . .
6The Problem
- BUT
- multimedia applications like
- movies
- high-quality videoconference
- Web/video/audio hybrid forms
- have stringent Quality of Service requirements
(data rate, latency, BER, etc)
7But . . .
- Current InterNet cannot guarantee QoS
- IP datagram based
- Too-late binding of resources to datagrams
8Dave Hudsons Nortel IdeaThe High Performance
Optical Overlay
- A second set of networks
- carries IP datagrams, but
- designed for guaranteed QoS
- augments the current InterNet
9The High Performance Optical Overlay
Admission controller
High Performance Optical Overlay
Client-side content switch
InterNet
clients
10Optical Overlay
- like Highway 401 North of Toronto
- free taxpayers pay collectivist
- no admission control
- no QoS versus
- Highway 407 Express Toll Route
- toll users pay free-market
- admission control
- QoS guarantee
11The High Performance Optical Overlay
- Observations
- 1 must contain some form of real circuit
construct fixed route, built before traffic
flows - else we cant bind resources to flows
- CANet 4, InterNet II architectures are,
happily, circuit-switched - 2 must limit resource consumption - e.g. by
session admission control - else we cant guarantee a session the resources
necessary to deliver specified QoS level
12Summary
- We seek Admission control algorithms which will
- guarantee QoS at some level to every admitted
session, and - admit upgrade sessions so as to optimize a
network utility function e.g. - constrained optimization problem
13The Utility Model
- Original Purpose manage the consumption of
resources in a multimedia server - Examples CPU time and bus bandwidths
- Each session has a Service Level Agreement
- a list of sets of resource requirements
- One set per level of QoS AND
- a value of utility assigned to each set ()
14QoS Levels
- Selection of a level of QoS
- (e. g. Gold, Silver or Bronze)
- implies
- value of utility to optimize
- (e.g. revenue in per hour) subject to
- set of resource requirements (constraints)
15 The Utility Model (UM) this famous figure
pinched by Jong-Wook Baek - IEEE Comm. Mag.,
March 01
session i
gold
bronze
silver
Quality Qi
session utility function
resource mapping
session utility ui(Qi)
session resources r (Qi)
system utility objective
system resource constraints
å r (Qi) R
16 Knapsack Problems (KP)
Volume p 39
17Multidimensional Multiconstraint Knapsack
Problem (MMKP)
18The Utility Model as an MMKP (Khans insight)
Let stones be sessions at QoS levels (Gold,
Silver, Bronze ) piles of stones be
sessions volume be multidimensional (multiple
resources) Choose one stone per pile to
maximize utility - usually revenue - (weight)
subject to all resource constraints
(multidimensional volume)
19Solving the Adaptive Media Problem (deciding
which sessions to admit at which QoS levels to
optimize utility)
20 Solving the MMKP
algorithm BBLP gives optimal solutions -
slowly heuristic HEU gives approximate but fast
solutions - promising for real-time SLA
admission control
21 But what about the High Performance Optical
Overlay??
22 - Apply the Utility Model
- to the links of a data network
- resources to manipulate
- bandwidth (conserved) and latency (not conserved)
of each link - (instead of server cpu cycles, I/O bw and Mp
bytes)
23Problems
- Determine which path a proposed new or changed
flow (e.g. an MPLS or RSVP flow ) should use,
based on - satisfaction of QoS constraints (feasibility)
i.e. - enough free bandwidth
- low enough latency on the path
- Optimal revenue utility
- Admission control decision - should the flow be
admitted ? (Optimal utility )
24Assumptions
- fixed latency ignore queuing
- simplifies things, can be relaxed
- adequate switch capacities
- ditto
-
25 Assumptions
- setup and teardown time for flows are
negligible - wont split a session over several paths
26 Assumptions
- Network supports fixed routing of flows, e.g.
RSVP/CoS, MPLS, Int II - (present InterNets best-effort datagram service
considered harmful ) - Auctions good, fixed pricing bad
27 Observation
- In a non-sparse network, the number of paths that
can satisfy the resource requirements for any
given flow may be VERY large - so bound the number of possibilities studied -
- heuristic, not algorithm
28Methodby an example
29An Example
N
30An Example QoS requirement increases
- Assume that Flow (S, D) has increased by ?
- Given
- Former state of the network N
- Old new traffic matrices T and T
- Old routing table R of N including (S, D)
- All old link capacities
31 - Need to find
- (possibly) new routing for (S,D)
- and maybe other flows
- New link capacities
- New state of N , including revenue-optimal
selection of (session, QoS_level) pairs
32Tactics
- 1 change nothing
- Existing route has enough surplus capacity? else
- 2 try to re-route (S,D) changing nothing else
(spare capacity elsewhere) , else - 3 try rerouting (S,D) and other flows (Benes)
, else - 4 increase link capacities somewhere ()
33To optimize revenue while respecting QoS
constraints
The Utility Model
34Gory Details . . .Or skip to 39
351 2 try to re-route (S.D), changing nothing
else
- Look for paths (S,D) with spare capacity
- 1 existing path (S,D)
T(S,D) T D
361 2 try to re-route (S.D), changing nothing
else
- 2 other paths
- 1 generate paths P(S,D) which are short enough
(acceptable latencies -- k shortest paths
algorithms)
P
D
shortest
shortest 1
shortest 2
S
Find k of them
371 2 try to re-route (S.D), changing nothing
else
- Invoke UM on P to
- 2 find
- CAP bw
CAP
D
381 2 try to re-route (S.D), changing nothing
else
- 3 select cap in CAP of maximum revenue
cap
D
Max
S
39How?
- Remove links by Shortest Paths Alg from N to
create P, - k shortest paths from S to D
- Apply Utility Model to
- P,
- Submatrix of the new traffic matrix,
- existing routing
- to get CAP and then cap
P
T(S,D)
The Utility Model
CAP
D
403 reroute existing paths
414 increase link capacities somewhere
- Homogeneous case assume cost of adding delta to
capacity of all links of N is constant
42increase link capacities somewhere
- 1 increase (S,D) by ?. Call UM. If revenue
increases, stop. Else - 2 increase (S,D)1 by ?. Call UM. If revenue
increases, stop. Else . . . - (continue until nearly out of time)
- N advise the customer to increase her bid by ?
- This is a heuristic!
43Experimental results
44Parameters
- 35-node real Enterprise network
- batches of O(100) SLAs
- Branch and bound algorithm used
- to determine optimal solutions in order
- to evaluate fast heuristics
- most runs on a slow Pentium II some on PIII
- Java implementation (perhaps X10 speedup feasible)
45Very preliminary results
46NB
- Contention typically begins at 10 - 15 SLAs
(using our SLA maker and 10000 unit bandwidths) - of Paths k 5 typically of little value
increases computation cost a lot - number of QOS levels / SLA 3
- Low contention scenarios give VERY good (optimal)
results in SLAOpt
47RuntimeNortel Topology20 paths /SLA are too
many!
48RuntimeNortel topology 5 paths/SLA is more
like it.
49Nortel topology revenue (utility)
50How close to maximum utility ?
- For
- no contention
- everybody gets admitted at Gold QoS
- SLAOpt Utility BBLP utility, as it should
- For 20 SLAs
- BBLP takes days of runtime, so no exact answer
yet - Data
51Performance of SLAOpt in 9 node Networkis 80-95
of optimal in this case
52Performance of SLAOpt in 31 node Nortel
NetworkPIII 800 Mhz is 85-90 of optimal
53Performance of SLAOpt in 100 node Network, PIII
800 MHzshowing the effect of k too large on
runtime ... K 3 or 5 is a bad idea
54Further research
- Explore the tuning of parameters of SLAOpt
- value of k in k shortest paths
- number of paths considered in k shortest paths
- tree pruning in searches
- value of tactic 3
- Carefully measure SLAOpt performance
55N.B.
- These runtimes of a few seconds (100 SLAs,
Nortel customer network) obtained from Java
implementation on 500 MHz Pentium - re-implementing in C will yield about 10x
speedup, according to local Java experts - fast enough for realtime admission with a
batching epoch of a second or more
56Implementation QoSNet
-
- prototype for High Performance Optical Overlay
- Tim Ducharme
57QoSNET Concept
Use Admission Controller (SLACtl) to admit
/reject traffic streams Customer requests
admission by specifying the stream parameters
(BW, latency,) using SLAs (Service Level
Agreements) SLACtl routes the new stream,
then determines if it can admit it while
respecting current SLAs,
58Admission Controller
Based on Khans Utility Model Adaptation of
Watsons SLAOpt simulator Shown theoretical
results of 80 resource utilization Communicates
with PassPort 7400 switches to setup/teardown
MPLS paths
59MPLS
Multi-Protocol Label Switching Need ER-MPLS
(Explicit Routed) to provide externally-specifiabl
e end-to-end routes through the network Alpha
version available only in latest release routing
software (pcr2.3 -mid June 2001)
60Alternative
- RSVP with Diffserv and COS
- unavailable from Nortel,
- available only in prototype from Cisco
- maybe less scalable?
61Test Process
62Recent research progress
- Mostofa Akbars work
- Distributed version of the admission controller
- Controls interconnected set of Enterprise
Networks ENs and servers - Modelled as interconnected set of MMKPS-
- The Multiple Knapsack MMKP or MMMKP
63Distributed SLAOpt
traffic
servers
64Approximate Solutions to the MMMKP?
- First try
- centralized algorithm calculations replicated by
complete negotiations among sites to find optimum - No loss of optimality naturally, but
- Message traffic far too high
65Approximate Solutions to the MMMKP?
- Second try
- Simplifying approximations
- Less iteration among nodes to search for global
optimum - Results about 70 of optimal
- No use of path tactics, heuristics purely at the
MMKP level - Tolerable message traffic
- Detailed evaluation going on now
66More ideas. . .
- Network operator accepts bids from lambda-vendors
- Lambda-vendors accept bids from fibre-vendors
- Fibre-vendors accept bids from conduit-vendors
- Multiple interacting auctions
- Can be modelled by array of MMKPs?
- Properties?
67THE END