1 smart networks project overview cisco: – david jaffe, karl auerbach, anna charny berkeley: –...
Post on 21-Dec-2015
217 views
TRANSCRIPT
1
Smart Networks Project Overview
• Cisco:
– David Jaffe, Karl Auerbach, Anna Charny
• Berkeley:– Venkat Anantharam, David Tse, Pravin Varaiya, Jean Walrand
– Stavros Tripakis (Post-Doc)
– Linhai He, John Musacchio, Jeff Danley, Jun Shu, Gaurav Agarwal, Eric Chi, Rajarshi Gupta
2
Outline
• Project Description– Why Smart Networks?
– Smart Networks Vision
– Main Focus Areas
– DiffServ
– Traffic engineering with MPLS/RSVP
– Test-bed
• Synergies with other groups
3
Why Smart Networks?
• Only the most basic measurements are used today for Networks Operations.
• Using more advanced measurements could improve– Planning
– Provisioning
– Resource Allocation to Different Classes of Service
– Fault Detection
• Need suitable methodology
4
Smart Networks Project
• DiffServ – Network Planning
– CoS Provisioning
– Operations: e.g., bandwidth brokers, tuning
• MPLS– Route-selection algorithms
– Robustness to link/node failures: selection of back-up routes
• DiffServ / MPLS integration– Traffic mixing at router’s scheduler
• Test-bed set-up for implementation, proof-of-concept and small-scale experimentation
Main Focus Areas
5
Smart Networks Vision
• Model networks, using knowledge from – Stochastic Models
– Applied Statistics, Simulation Techniques
– Control Systems
• Measurements -> Model Fitting -> Analysis/Simulation -> Decision Making
DesignerOperator
Historical
RecordsModel Fitting and Validation
(Model Selection)
Model Evaluation and Optimization(Analysis & Simulation)
OAM Message Transport
Probes; MIB
Design, Planning, Operations
Network Devices
Tools
6
DiffServ• Goal:
– CoS without per-connection state
• Solution: – No route-pinning (routing protocol, e.g. OSPF, BGP, handles that)– Planning and operations based on aggregate statistics and worst-case
routing– Admission-control, policing/shaping, tagging at the edge– Peer-to-peer SLAs that specify total rate but not traffic destination
SLACloud 1 Cloud 2
Edge router Core router/switch
7
DiffServ Issues
• What is the right admission control policy:– Centralized/Distributed– Depends on client request model: pipe/hose
• Pipe model: client requests access to specific destination (i.e., ingress and egress routers are specified)
• Hose: client requests access to the network (destination of client unknown, only ingress router known)
– Worst-case / Statistics based (measuring temporal and spatial distribution of traffic)
8
DiffServ Issues (continued)
• How are SLAs updated (resizing) :– Peer-to-peer: bandwidth broker only talks to peer
clouds. Algorithms/protocols ?• Convergence of updates, stability• Performance w.r.t. centralized solution
– End-to-end: request travels end-to-end and updates SLAs along the path(s).
• Scalability ? (Hierarchical solutions ?)• Acceptance of hierarchical solutions by cloud administrators ?
9
DiffServ Issues (continued)
• Analysis to establish end-to-end guarantees:– What is the right model for a router ?
• Highly dependent on scheduling algorithm, but would like to abstract
• Per-hop-behavior (PHB) definitions not formal/strict enough
– What are the right assumptions for intra- and inter-domain traffic statistics ?
– How does best-effort (or AF) traffic affect EF traffic ?
• On-line monitoring and reaction to failures:– Is re-routing (handled by routing protocol) enough or is rescheduling at routers also necessary ?
10
Ingress 1
Ingress 2
Ingress 3
Typical Case(spatial distribution)
Ingress 1
Ingress 3
Worst Case
Bottleneck LinkIngress 2
DiffServ : Worst Case Admission Control
Admission control based on worst case is excessively wasteful!
11
DiffServ : Measurement-Based Admission Control
New
Capacity
Mean + 2.4
Admit if peak(new) < Gap at all times
Gap
12
DiffServ : Measurement-Based SLA resizing
Use same idea for SLA resizing
• Keep statistics of traffic flowing to peer clouds• Request SLA update using:
• damping: update only when capacity of SLA exceeded• over-provisioning: update to more than actually needed
• Problem: what happens if peer broker rejects update request ?
13
Traffic Engineering with MPLS/RSVP
• Goal : guarantee QOS by explicitly controlling the routes
• How ?– MPLS (multi-protocol label switching) :
• Like ATM virtual-circuit switching• Packets encapsulated with a label (tag)• Router forwards based on label/input port, and assigns new label• Allows to bypass the IP routing protocol
– RSVP (resource reservation protocol) :• Signaling protocol, to establish the MPLS path initially• Also does resource book-keeping: how many paths are using a link
and how much of the capacity of that link is available• Soft state: reservations are periodically refreshed, to account for
failures, etc
Path Setup - Example
R8
R2
R6
R3
R4
R7
R1R5
R9
Setup: Path (R1->R2->R6->R7->R4->R9)Setup: Path (R1->R2->R6->R7->R4->R9)
Reply: Resv communicates labels andReply: Resv communicates labels andreserves bandwidth on each linkreserves bandwidth on each link
Pop
22
4917
32
15
Traffic Engineering Issues
• Off-line:– Plan routes for all possible source/destination pairs and
network states ahead of time (“optimal” solution).– Need knowledge of traffic statistics (arrival rates, holding
times, requested bandwidth) but these requests are not just voice calls ! (Might be aggregates coming from another cloud.)
• On-line:– Route requests as they arrive– Not using history/prediction of traffic => “greedy” approach
What is the right policy for path selection?
16
Traffic Engineering Issues
• Centralized/Distributed ?– Centralized assumes all requests arrive at the same route-selection
module.– Distributed may result in inconsistencies (e.g., over-booking).
• Many alternatives, even for on-line, centralized : – Any min-hop path with available resources– “Widest” path– Widest min-hop path– Minimum-interference path– Some weighted combination of some of the above ?
17
Traffic Engineering Issues (continued)
• How to make it robust to failures (e.g., single link/node failure) :– Reserve two (disjoint) paths for each request : active and back-up– Difficult choice of pairs of active / back-up paths
• How to periodically re-optimize by re-routing :– Example:might need to re-route a number of small-size calls to
accommodate / balance a large-size call.
• How to choose paths w.r.t. multiple-criteria :– Example: find a path that does not traverse a particular type of router
more than k times.
• End-to-end issues :– “Optimal” routes per cloud might not form an end-to-end optimal route.
18
DiffServ and Traffic EngineeringCo-existence
• Technologies are essentially complementary• Use DiffServ for aggregation into few traffic classes• Use traffic engineering to by-pass IP routing
• Issues for MPLS and DiffServ integration:– Path establishment in MPLS does not imply resource allocation
guarantees : need analytical techniques to establish end-to-end delay bounds.
– Scheduling mixed traffic:• Tag MPLS packets with DiffServ PHB (per-hop-behavior) tags ?• Hierarchical scheduling ?
20
Smart Nets Test Bed - San Jose
Legend
10MbpsEthernet
21
Management and Control Network
Legend
Serial(connected to
router’s console)
22
UNIX Box
• Receives Measurements from Routers (local or end-to-end measurements) :– Available Bandwidth
– Loss Rate
– Delay
– Jitter
23
UNIX Box
• Uses Measurements to Control Router– Adjusts Router’s Scheduling configuration, e.g.:
• Priority or Weighted Fair Queuing
• Adjust priorities or weights
– Currently building API using Expect (Tcl similar):
Jeff Danley, UCB student.
– Automated Process
24
Synergies
• Models / data for wide-area traffic
• Simulation tools (e.g., SSFNET, NIMI, for WAN)
• Test-bed:– Traffic generation software
– Measurement software