resource allocation and queuing theory -...
TRANSCRIPT
IntroductionResource Allocation
Resource Allocation and Queuing TheoryModeling
Joseph Spring
Modeling Networks
Joseph Spring Queuing Theory
IntroductionResource Allocation
Outline
1 IntroductionWhy are we waiting? . . .
2 Resource AllocationIssuesNetwork Models
Packet-Switched NetworkConnectionless FlowsService Model
TaxonomyRouter-Centric versus Host-CentricReservation Based versus Feedback-BasedWindow-Based versus Rate-BasedWindow-Based versus Rate-Based
Joseph Spring Queuing Theory
IntroductionResource Allocation
Why are we waiting? . . .
Outline
1 IntroductionWhy are we waiting? . . .
2 Resource AllocationIssuesNetwork Models
Packet-Switched NetworkConnectionless FlowsService Model
TaxonomyRouter-Centric versus Host-CentricReservation Based versus Feedback-BasedWindow-Based versus Rate-BasedWindow-Based versus Rate-Based
Joseph Spring Queuing Theory
IntroductionResource Allocation
Why are we waiting? . . .
Why are we waiting?Fundamentals
There are lots of examples for queuing situations:
Passport control
Supermarkets
Planes waiting to take off or land
Programs being processed by uniprocessor/multiprocessor
Packets at sending side of communication channel
Packets at receiving side of communication channel
Each of the above involves different states: waiting, beingprocessed, processed, together with transitions between eachof the states.
Joseph Spring Queuing Theory
IntroductionResource Allocation
Why are we waiting? . . .
Why are we waiting?Fundamentals
Waiting is a consequence of the randomness of the processingprocedure
arrival times - not known in advance
processing times - not known in advance
number of times we have to wait in a queue for either thesame or a different service - not known in advance
If these were known then we could schedule in advance andeliminate waiting altogether.
In order to model queuing scenarios
we employ stochastic processes reflecting the randomindeterminate nature of the particular situation beingstudied.
Joseph Spring Queuing Theory
IntroductionResource Allocation
Why are we waiting? . . .
Why are we waiting?How many more times . . . ?
We explore operations of interest operating under randomprocesses in order to gain a better understanding of theoperations characteristics. We might ask
how long do we expect to wait prior to processing? -(customer question)
how long is a service likly to be idle? - (service question)
how can a reasonable balance be obtained between thetwo questions above?
We seek to identify those characteristics of a system that willallow us to measure the performance of a system.
Joseph Spring Queuing Theory
IntroductionResource Allocation
Why are we waiting? . . .
Why are we waiting?Fundamentals
In working with queues we need to acquaint ourselves with:AssumptionsActorsArrival and Service Times - (Modeled by StochasticProcesses)Service Discipline - (FIFO, FQ, LIFO, SIRO, Priorities)Design and Execution - (Single, Parallel, Series, Tandem,Network)Queue size - (finite, infinite)nature of Calling Source - (generates finite/infinite numberof customers)Customer and service behaviour - (processing speed afunction of length of queue; customer: jockeying, balking,reneging)
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Outline
1 IntroductionWhy are we waiting? . . .
2 Resource AllocationIssuesNetwork Models
Packet-Switched NetworkConnectionless FlowsService Model
TaxonomyRouter-Centric versus Host-CentricReservation Based versus Feedback-BasedWindow-Based versus Rate-BasedWindow-Based versus Rate-Based
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Resource AllocationIssues
Resource Allocation and Congestion Control arelongstanding active areas of researchResource allocation partially implemented in
routers / switches in a networkthe transport protocol running on the end hosts
End systems use signalling protocols to indicate resourcerequirements to network nodes
Network nodes respond with information on resourceavailability
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Resource AllocationIssues
Definition
Resource Allocation is defined to be the process by whichnetwork elements try to meet the competing demands thatapplications have for network resources, by which we mean
Link bandwidth (BW)
Buffer space in routers / switches
Note:
it is not always possible to meet all of the demands placedon a resource
some users / applications may receive fewer resourcesthan they require
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Resource AllocationIssues
Throughout the following discussion
flow control will refer to practices that ensure a fast senderdoesn’t overpower a slow receiver
congestion control will refer to practices that ensure a setof senders do not send too much data into the network dueto a lack of resources at some point
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Outline
1 IntroductionWhy are we waiting? . . .
2 Resource AllocationIssuesNetwork Models
Packet-Switched NetworkConnectionless FlowsService Model
TaxonomyRouter-Centric versus Host-CentricReservation Based versus Feedback-BasedWindow-Based versus Rate-BasedWindow-Based versus Rate-Based
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Resource AllocationNetwork Models
Packet Switched NetworkMultiple links and switchespackets may encounter a bottleneck at some point in thenetwork in which, for example high speed links feed into alow speed link
Figure: A Potential Bottleneck Router
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Resource AllocationNetwork Models
Figure: A Potential Bottleneck Router
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Resource AllocationNetwork Models - Connectionless Flows
Assume network essentially connectionless anyconnection oriented service implemented in transportprotocol running on end hostsInternet model
Sequence of packets (datagram delivery service)TCP implements end to end connection abstraction
connection set up message traverses networkreserves a set of buffers for connection at routers (underutilisation of available resources since not available forgeneral use)
we focus on connectionless networks
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Resource AllocationNetwork Models - Connectionless Flows
What do we mean by connectionless?All datagrams are completely independent - Modeled byPoisson Distribution - car modelAlthough datagrams are switched independently usually thecase that stream of datagrams flow between a particularpair of hosts through a particular pair of hosts - train model
Definition
Model assumes a flow abstraction as a sequence of packetsbetween source and destination following the same routethrough the network
flows can be defined at different granularitiesHost to host (same source / destination addresses)process to process (same source / destination host / portpairs
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Resource AllocationNetwork Models - Connectionless Flows
Figure: Multiple flows passing through a set of routers
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Resource AllocationNetwork Models - Connectionless Flows
State information
for each flow kept at routers (soft state)
used to make resource allocation decisions regardingpackets that belong to the flow
Soft state
not always explicitly created and removed by signallinga middle ground between
purely connectionless network maintaining no state atrouterspurely connection oriented network maintaining hard stateat routers
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Resource AllocationNetwork Models - Connectionless Flows
Soft state
exists only for purpose of resource allocation
not essential for correct operation of network
however router better able to handle packets belonging toflow if soft state maintained for flow
does not imply relible and ordered delivery of a virtualcircuit
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Resource AllocationNetwork Models - Service Model
In the Best Effort Service Model
each packet is treated in the same way
flows are given no guarantees
Defining a service that gives some form of guarantees
for example the bandwidth required for a video stream
is said to provide multiple qualities of service (QoS).Clearly a range of possibilities exist, regarding the optionsavailable here.
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Outline
1 IntroductionWhy are we waiting? . . .
2 Resource AllocationIssuesNetwork Models
Packet-Switched NetworkConnectionless FlowsService Model
TaxonomyRouter-Centric versus Host-CentricReservation Based versus Feedback-BasedWindow-Based versus Rate-BasedWindow-Based versus Rate-Based
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Resource AllocationTaxonomy - Router-Centric versus Host-Centric
Two broad groupsThose that address the problem of Resource allocationfrom within the network (routers / switches)Those that address the problem from the edges (end hosts)
Real issue is considered to be where does the mainburden resides
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Resource AllocationTaxonomy - Router-Centric versus Host-Centric
Router-centriceach router takes responsibility for deciding when packetsare forwardedwhich packets are droppedinforming hosts generating traffic how many packets theyare allowed to send
Host-centricend hosts observe network conditions (how many packetsare getting through) and adjust accordingly
We note that these groups are not mutually exclusive, bothplay their part
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Resource AllocationTaxonomy - Reservation Based versus Feedback-Based
Should we make Reservations or wait for Feedback?Reservation Based System
End hosts ask for network for a certain amount of capacityat time when flow establishedEach router allocates resources (buffers and/or % of linkbandwidth)If router cannot commit to requirements then requestdenied
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Resource AllocationTaxonomy - Reservation Based versus Feedback-Based
Feedback Based SystemEnd hosts send data without recourse to networkEnd Hosts adjust sending rate subject to feedback
EXPLICIT feedback - Router sends message ’slow down’IMPLICIT feedback - end host adjusts sending rate subjectto network conditions e.g. packet loss
Reservation Based implies router centric resourceallocation since each router responsible for allocation andpolicing of its capacityExplicit Feedback-Based implies router centric basedmechanismImplicit Feedback-Based implies Host centric mechanism,routers ’silently drop packets when congested
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Resource AllocationTaxonomy - Window-Based versus Rate-Based
Two general ways to express to the sender the amount ofdata that may be sent
Windows based transport protocol indicates the size ofwindow (amount of buffer space) available at the receiversendThis limits amount of data that sender can send - so a flowcontrol mechanisma similar mechanism window advertisement can be used toreserve buffer space - so a reource allocation mechanism(e.g. X.25)
Joseph Spring Queuing Theory
IntroductionResource Allocation
IssuesNetwork ModelsTaxonomy
Resource AllocationTaxonomy - Window-Based versus Rate-Based
Rate basedA Rate based transport protocol indicates the size ofindicates the amount of data that the receiver or network isable to absorbA rate based characterisation of flow is felt to be a logicalchoice in a reservation based system supporting differentqualities of serviceSender makes reservation for so many bits per second,each router on path determines if it can commit toreservation subject to current committments
See summary on page 464
Joseph Spring Queuing Theory