Globally Optimal Distributed BatchReconfiguration for
Hazard-free Dynamic Provisioning:How an Entire Network can Think
Globally and Act LocallyWayne D. [email protected]
of Alberta and TRLabsEdmonton, AB, Canada
DRCN 2007 La Rochelle, France, Oct. 70-10
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
Setting the stage..what motivates this proposal?US National
Science Foundation:Calls for completely new approaches to network
operations.Identifies robust networking as one of the grand
challenges in networking science Concern that existing peer-to-peer
asynchronous distributed provisioning scheme has the risks of
network state incoherenceE.g. [1] Pandi & Wosinska,
ICTON-RONEXT 2005Separately, in other industries, there is a move
to exploring the applications and benefits of on-line O.R. Existing
distributed provisioning schemes can only employ greedy solution
methodsWhat if a whole network could think globally, but act
locally?Greater resource efficiencies, greatly reduced signalling,
hazard-free operation, continual near-optimality (self
consolidating)
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
What Problem(s) are we trying to solve in dynamic protected
service provisioning?The inherent risk of schemes that operate
dynamically, on the per-connection timescale, assuming global state
coherency at all times. Risky ! High signalling volumesRather than
trying to quantity and lower the risk: is there some approach that
fundamentally avoids the risk in the first place?In existing
concepts provisioning is per-path with no chance to globally
optimize Periodic re-optimization of overall network configuration
is awkward or unaddressed.
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
OverviewWhat is the problem?Review key aspects of current
dynamic provisioning concept Key Concepts of New ProposalOutline of
OperationSub-study: Benefits of Batch Incremental Re-optimization
problemSample resultsSummary of Advantages and
DisadvantagesResearch Directions
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
SBPP Dynamic Protected Service Provisioning Concept(1)
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
SBPP route computation and signaling process1. Compute working
and protection routesWorking: 0-4-8-11; Protection: 0-3-7-9-12-112.
Establish working path3. Establish protection pathSBPP Dynamic
Protected Service Provisioning Concept(2)
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
Observations / Concerns about Dynamic SBPP Every node needs and
assumes a complete and current network state database, and existing
current protection capacity sharing relationships Link state
updates are advertised on a per-connection basis Link state updates
are disseminated asynchronously by any node at the same time other
nodes are relying and acting upon time critical state information.
The total database of network state that is operationally critical
grows at least as O(n3) with size of the network or operating
domain and also intensifies with frequency of changes in the
network.
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
Alternatives for Dynamic Automated Provisioning Centralized
Control: Global view, one operation at a time.Safe (in the present
regard) but other downsides Apply packet priorities to update
messages, use TE summary packets, etc. i.e., measures to try to
just mitigate the riskWill eventually crash when provisioning is
dynamic enoughProtected Working Capacity Envelope ConceptRemoves
protection arrangements from the per-connection time scale Refs:
Grover- Comm Mag, Shen & Grover, Shen PhD; available at
www.ece.ualberta.ca/~groverProposal: Globally Optimal Distributed
Synchronous Batch Re-optimization Eliminates the hazard of database
incoherenceFramework yields other advantages
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
Key Concepts of New Proposal Nodes in these networks have
precise time ! Can we exploit that?YES: Time synchronization can
help in data synchronization Small-batch incremental reoptimization
provisioning not path-by-path instantaneous asynchronous
provisioningGlobally synchronous change actions, not asynchronous
actionsReliance on precise time to coordinate actions and
decisions.Relegating all operationally critical signaling for state
update to non- real-time communication requirementsRobust
confirmation of global state database coherence before any reliance
upon it for network actionsSolving a globally optimal
reconfiguration solution But nodes act locally to put into effect
their parts only of globally optimal reconfiguration plans.
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
How it works:Operational phases:
Batch Change AccumulationChange Dissemination and
ConfirmationGlobally Optimal Reconfiguration SolutionLocal change
activation
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
Step 1. Batch Change Accumulation5 to 10 minute interval
envisagedMore generally, the period is relative to the connection
holding time and request rate (I.e. provisioning traffic
intensity)Nodes make no changes to network connection state during
this timeNodes observe:New requestsDepartures (released
connections)as they arise at their location only.At end of the
period nodes emit a change summary packetLike an existing LSA, but
contains batch change infoRobust error detection / correction
encoded on packetThis dissemination is not real-time-critical
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
Step 2. Change Dissemination and Confirmation(Again, no change
is made to network state made in this phase)Nodes receive batch
change summary packets from each other.SLA-like forwarding as in
OSPF (Internet)May include pre-arranged scheduled service
requestsThis data exchange is not real-time criticalThis process
overlaps with the next change Accumulation phaseNodes integrate
change packets received into single network-wide re-provisioning
summary view of the new requirements.Each node then emits a global
change summary checksumEach node wait until an intermediate time
mark: If every heard checksum matches own: proceed,Else: flood out
wave-off: go around signal
Partial change listPartial change listPartial change
listChecksum of integrated overall network change listChecksum of
integrated overall network change list
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
Step 3. Thinking Globally: Optimal Reconfiguration Solution
Each node locally solves an instance of the globally optimal
reconfiguration problemMay be any problem version network operator
prefersExample: Route and protect maximum number of the new service
requestsWhile reclaiming capacity from released connectionsWith or
without permission to re-optimize existing protectionother
Variants: Multiple priorities or protection classes
(multi-QoP)Permission to re-arrange selected working
pathsStrategies to include hedging against future
uncertaintyImpairment aware, availability aware routing, etc.Nodal
solutions have to be identical not just equivalentProspect here for
true on-line O.R.any reduced complexity version of the optimal
problem can also be substituted here
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
Step 4.Acting Locally: Node do their part of the solution
(only)
On the next globally precise-time mark:Each node activates the
switching matrix changes to put into effect its part (only) of the
complete network reconfiguration solution.No continuing existing
connection is alteredService access nodes observe the turn-up of
their new connections and test end-to-end.New operating phase
commencesChange request accumulation continuesNote that this
results in creation of a complete set of new service paths and
their protection arrangements simultaneously in parallel on the
network with no signaling. Correctness of the outcome is
independently validated by each end-node pair (as it would be in
any case).
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
Overall Network Synchronous Phases
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
Sub-Study: Benefits of Optimal Batch Incremental Re-Provisioning
(with Z. Pandi on COST 270 STSM to TRLabs)
Simulation of an on-line O.R. application for batch incremental
re-optimization that is made possible by this framework.
Statistically non-stationary random traffic demand i.e., not
just random but spatially and temporally evolving random arrival /
departure trafficTests / illustrates ability for scheme to
inherently track and re-optimize for time-evolving demand patterns
Each node accumulates batch change infoAt end of each batch period,
globally optimal incremental reconfiguration problem solved (on a
single CPU)Global changes put into effect locally in simulated
networkCompared performance against asynchronous independent
provisioning using best known SBPP provision algorithm
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
Simulation Details
SBPP protection principle vs. small batch incremental
reoptimization. (AMPL Model is Appendix to the paper)Spare capacity
allocations re-optimized each interval as well as new and released
working paths routedNetworksSparse topology High degree topology
ScenariosStationary randomTemporal overloadTemporo-spatial N-S and
E-W evolutions; Accumulation intervals from 0.2 to 0.4 holding
times
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
Test Networks
Original EU ModelSparse VersionTime-space non-stationary
statistical evolution of demand pattern
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
Sample Performance ResultsFull topology, general overloadFull
topology, spatial evolutionSparse topology, spatial evolutionTime
(unit holding times)Total number of blocking eventsBatching
interval = 0.4 mean holding time
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
Summary: Properties, Advantages, Disadvantages
(+) Eliminates the hazard of database incoherency under
asynchronous operation All critical state-exchange becomes non-time
critical (+)(+) Network enjoys the efficiency and adaptability of
on-line continual global re-optimization of network state(-?) New
connections are provisioned in the next provisioning cycle, not
instantaneously.( service activation delay)Nodes synchronize their
actions using existing network network time/frequency assets.
analogy to clocked logic circuit robustnessService protection still
acts at any time in response to actual failure, provisioning cycle
skipped so protection action is reflected in next change
accumulation period.
Wayne D. GroverUniversity of Alberta and TRLabs *Globally
Optimal Distributed Synchronous Batch Reconfiguration
Research Directions within this framework
Incremental re-optimization models and strategies that the
framework enables Options such as spare capacity re-optimization or
notMulti QoP classes, prioritiesWorking re-arrangeable service
classes?Maximum revenue, minimum load, etc: different objectives
Multi-QoP provisioning solutionsIncremental on-line grooming
optimizationDifferent approaches to identical not just equivalent
solution of the disparate instances of the same global optimization
problem.Links to the scheduled lightpath connection planning
problem and the network consolidation problem.Accommodation for a
top-priority no-delay service classIf thought essentialExtension to
Domains rather than nodesLinks to the PWCE conceptCollaborations in
this area already begun with B. Jaumard, Networks OR group at
Concordia U., Montreal
