Download presentation
Presentation is loading. Please wait.
2
UCB
3
Tools for Smart Networks Jean Walrand BITS (Berkeley Information Technology & Systems) U.C. Berkeley www.eecs.berkeley.edu/~wlr/mascots2000
4
UCB Outline What are Smart Networks? Why Smart Networks? Tools for Smart Networks Project Example 1: DiffServ Example 2: Bandwidth Allocation Example 3: Server Allocation Conclusions
5
UCB What are Smart Networks? Measure Analyze Modify
6
UCB Why Smart Networks? Before: “Simple Network” IP Transport Applications Transport Applications Client Server Network
7
UCB Why Smart Networks? (continued) Now: “Complex Network” Network Application Servers, Content Servers Caches, Traffic Shapers, Redirection Agents, Processing
8
UCB Why Smart Networks? (continued) Simple Network: IP Forwarding Routing Table Updates DNS Intelligence in Hosts Complex Network: New Functions New Transport Services (e.g., CoS, SLAs) Needs Intelligence in Network
9
UCB Why Smart Networks? (continued) INTERNET IP LANs, ATM,... Applications Success of SimplicitySuccess of Complexity TELEPHONE SS7, Billing,... OC-n, DS-n, UTP Applications
10
UCB Why Smart Networks? (continued) Probably not very desirable! INTERNET IP LANs, ATM,... Applications
11
UCB Why Smart Networks? (continued) INTERNET IP LANs, ATM,... Applications IP LANs, ATM,... Applications M/A/M Tools for Planning, Design, Operations
12
UCB Tools for Smart Networks Project Joint UCB - Cisco Project DARPA Funding + Cisco Combines Measurements Analysis & Simulation Real-time Control Objective: Product
13
UCB Tools for Smart Networks Project (cd) Utility Comprehension Simulations Measurements Analysis Integrated Tools
14
UCB Tools for Smart Networks Project (cd) Cisco: David Jaffe (Lead Investigator) Karl Auerbach (Lab Design and Implementation) Anna Charny (MPLS) TBS (DiffServ) UCB Anantharam, Tse, Varaiya, Walrand Stavros Tripakis (post-doctoral scholar) About 6 graduate students TEAM:
15
UCB Example 1: DiffServ Goal: CoS without per-connection state No route-pinning Planning and operations based on aggregate statistics and worst-case routing Peer-to-peer SLAs that specify total rate but not traffic destination Solution:
16
UCB Example 1: DiffServ (continued) Cloud 2 Cloud 1 Policing Shaping SLA
17
UCB Example 1: DiffServ (continued) Ingress 1 Ingress 3 Ingress 1 Ingress 2 Ingress 3 Typical Case Worst Case Ingress 2 Bottleneck Link DiffServ SLA: Worst Case Admission Control Terribly wasteful!
18
UCB Example 1: DiffServ (continued) New Admit if peak(new) < Gap at all times Capacity Mean + 2.4s Gap DiffServ SLA: Measurement-Based Admission Control
19
UCB Example 1: DiffServ (continued) How well does this approach work? Simulation study: Construct traffic model (parametric FBM) Validate model against measurements Simulate admission control policy Test fraction of SLAs that see congested links and level of congestion Experimental study (coming year) Implement measurements and admission control Evaluate performance Work of Linhai He and John Musacchio
20
UCB Example 2: Bandwidth Allocation Problems: How to share bandwidth How to renegotiate SLAs Issues: Scalability Efficiency Fairness, Optimality,...
21
UCB Example 2: Bandwidth Allocation (cd) Sharing one link: N X Y Number of “calls” [Voice over IP]
22
UCB Example 2: Bandwidth Allocation (cd) N X Y Dynamic X Y N N Feasible region
23
UCB Example 2: Bandwidth Allocation (cd) N X Y Static N1N1 N2N2 Admission policies SLAs (Committed Access Rates) X Y N N N1N1 N2N2 Feasible region
24
UCB Example 2: Bandwidth Allocation (cd) X Y N N Dynamic N1N1 N2N2 Static
25
UCB Example 2: Bandwidth Allocation (cd) Closer Look: Assume Poisson demands, i.i.d. holding times... X Y N N For “large links”, the variance is small. => Static Dynamic However, rates change => must adapt
26
UCB Example 2: Bandwidth Allocation (cd) Proposed Adaptation Scheme: Renegotiate “blocks” of permits based on thresholds 10 35 3 4 6 4 15 567 5 67 20 40
27
UCB Example 2: Bandwidth Allocation (cd) How well does this approach work? Simulation study: Birth/Death Model of Bandwidth Study Efficiency vs. Rate of Renegotiation Work of Eric Chi and Linhai He
28
UCB Example 3: Server Allocation S S S S Location + Load
29
UCB Example 3: Server Allocation (continued) S S S S * * Anycast: Closest * Least Loaded Among N Closest Stats
30
UCB Example 3: Server Allocation (continued) Model: Conflict between measure lengths (by sending jobs to all queues) send only to queue believed to be shortest
31
UCB Example 3: Server Allocation (continued) Algorithm: send to queue k with probability f k (T 1,..., T K ) Example: f k (T 1,..., T K ) = (1/T k )/ (1/T 1 +... + 1/T K ) Not very sensitive to choice of function f k
32
UCB Example 3: Server Allocation (continued) Work of Gaurav Agarwal and Rahul Shah How well does this approach work? Simulation study: Construct traffic model (Poisson requests, random lengths) Simulate server allocation policy (ns) Test response times and server utilization
33
UCB Conclusions Common View: ResearchDevelopmentR&D Stochastic Models Performance Evaluation Limit Theorems.... Prototype Hacking Tuning.... AcademiaIndustry Bell Labs XEROX PARC...
34
UCB Conclusions More Accurate View: Academia Industry Comprehension-driven research Utility-driven research
35
UCB Thank You!
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.