97-1086R1: Per-VC Rate Allocation Techniques For ABR ...jain/atmf/ftp/af_vsv2.pdf · The Ohio State University Raj Jain 1 97-1086R1: Per-VC Rate Allocation Techniques For ABR Feedback

Raj JainThe Ohio State University

1

97-1086R1: Per-VC Rate97-1086R1: Per-VC RateAllocation TechniquesAllocation TechniquesFor ABR Feedback inFor ABR Feedback in

VS/VD NetworksVS/VD Networks

Rohit Goyal, Xiangrong Cai, Raj Jain, Sonia Fahmy,Bobby Vandalore

The Ohio State University

Contact: [email protected]

http://www.cis.ohio-state.edu/~jain/

Raj Jain

Horizontal small


2

Virtual Source / VirtualVirtual Source / VirtualDestination (VS / VD)Destination (VS / VD)

q Segments the end-to-end ABR control loop.

q Coupling between loops is implementation specific.

q VS/VD can help in buffer management across thenetwork.

q ABR switches separated by non-ATM networkcould also implement VS/VD.


3

GoalsGoalsq Describe a VS/VD switch architecture.

q Discuss issues in designing rate

allocation schemes for VS/VD switches.

q Present a per-VC rate allocation scheme for VS/VD.

q Discuss how VS/VD can help in buffer managementacross the network.


4

VS/VD SwitchVS/VD SwitchArchitectureArchitecture

SwitchingFabric

Data flowRM cell flow

Class Queue

Per-VC Queues

Per-VCScheduling

Port 1 Port 2

Otherports

Otherports


5

VS/VD SwitchVS/VD SwitchArchitectureArchitecture

q Each switch port :m Class queue for each service category.(optional)m Per-VC queues drain into class queue or link

q When a cell is received :m Data cell : forwarded to destination port (VS).m FRM cell : turned around as BRM (VD).m BRM cell : ER is noted (VS).

q VS sends data + FRM cells at ACR to class queue.q A scheduler services the per-VC queues.


6

A Simple VS/VD ModelA Simple VS/VD Model

Node i Ri

sijrij

ERijexternalERij

internalERijfeedback

qiqij

q Internal Service Rate = f(External/DownstreamFeedback, Local congestion)

q Local Congestion = f(Qi); Qi = qi + Σ qij

q Upstream feedback = Internal service rate

q Example: Downstream = 100 Mbps,Internal =90 Mbps = Upstream Feedback


7

Simple VS/VD ModelSimple VS/VD Model

q Desired input rate to class queue is also fed back tothe upstream switch.

q Problem:

m Transient per-VC queues cannot drain.Input rate sij = Output rate rij

m Queues that build up during open loop phase do notdrain.


8

Ri

sij

rij

ERijexternalERij

internalERijfeedback

qi

qij

Correct VS/VD ModelCorrect VS/VD Model

q Internal Service Rate = f(External/DownstreamFeedback, Switch algorithm using qi)

q ACRij = f(Internal service rate, end system rules)

q Upstream feedback =f(qij)ACRij

q Example: Downstream = 100, Service =90, ACR=80,Upstream feedback=70 Mbps


9

Per-VC ERICA+Per-VC ERICA+q BRM received :

m ERijexternal := ER in RM cell

q FRM received :

m ER in RM := ERijfeedback

q At the end of each averaging interval :

m ERijinternal

:= Min{ Max (rij/Overload, g(qi)Ri/N), ERijexternal}

m Output rateACRij = rij := fn{ERij

internal, end system rules}

m ERijfeedback := g(qij)rij


10

Simulation ModelSimulation Model

Switch 1

Switch 2

Switch 3

D1

D5

S1

S5

275 or 50 ms 5 ms 5 µs5 ms

45 Mbps155 Mbps155 Mbps 155 Mbps

Link 2Link 1

VS/VD Loops


11

ParametersParametersq VS/VD and Non-VS/VD configurations.

q First hop = Satellite hop with 1 way delay:

m LEO = 50 ms

m GEO = 275 ms

q Link 2 = 45 Mbps (Bottleneck Link).

q All other links = 155.52 Mbps (149.76 with SONET)

q Persistent ABR sources: ICR = 30 Mbps

q Persistent TCP sources: Timer granularity = 500 ms.At 45 Mbps, 100 ms causes timeouts in GEO.Known problem with TCP Std deviation measurement.


12

ERICA+ Non-VS/VDERICA+ Non-VS/VDLEOLEO


13

0

5000

10000

15000

20000

25000

30000

35000

0 10002000

30004000

50006000

70008000

900010000

Qu

eue L

eng

th

Time in milliseconds

Five ABR : SW2 Queue Length

120 ms × 100 Mbps = 19 k cells⇒ Q < 3 RTT

ERICA+ Non-VS/VDERICA+ Non-VS/VDLEOLEO


14

ERICA+ Non-VS/VDERICA+ Non-VS/VDGEOGEO


15

ERICA+ Non-VS/VDERICA+ Non-VS/VDGEOGEO

0

20000

40000

60000

80000

100000

120000

140000

160000

0 10002000

30004000

50006000

70008000

900010000

Qu

eue L

eng

th



570 ms × 100 Mbps = 134 kCells⇒ Q < 3 RTT


16

0

20000

40000

60000

80000

100000

120000

140000

160000

0 10002000

30004000

50006000

70008000

900010000

Qu

eue L

eng

th



Queues move from bottleneck hop to non-bottleneck hop but do not come down.

ERICA w/ Simple VS/VDERICA w/ Simple VS/VD


17

0

500

1000

1500

2000

2500

3000

0 10002000

30004000

50006000

7000

8000

9000

10000

Qu

eue L

eng

th


10 ms × 100 Mbps = 2.4 kCellsQ = Previous loop delay

ERICA w/ Simple VS/VDERICA w/ Simple VS/VDSw2 Queue


18

ERICA+ VS/VD LEOERICA+ VS/VD LEO

0

5000

10000

15000

20000

25000

30000

35000

0 10002000

30004000

50006000

70008000

900010000

Qu

eue L

eng

th


Five ABR VS/VD : SW1 Queue Length

100 ms × 100 Mbps = 24 kCellsQ ≈ Previous loop delay


19

ERICA+ VS/VD LEOERICA+ VS/VD LEO

0

500

1000

1500

2000

2500

3000

0 10002000

30004000

50006000

70008000

900010000

Qu

eue L

eng

th



10 ms × 100 Mbps = 2.4 kCellsQ ≈ Previous loop delay


20

VS/VD GEO Sw1 QueueVS/VD GEO Sw1 Queue

0

20000

40000

60000

80000

100000

120000

140000

160000

0 10002000

30004000

50006000

70008000

900010000

Qu

eue L

eng

th



550 ms × 100 Mbps = 130 kCellsQ ≈ Previous loop delay


21

0

500

1000

1500

2000

2500

3000

0 10002000

30004000

50006000

70008000

900010000

Qu

eue L

eng

th



10 ms × 100 Mbps = 2.4 kCellsQ ≈ Previous loop delay

ERICA+ VS/VD GEOERICA+ VS/VD GEO


22

ERICA+ VS/VD GEOERICA+ VS/VD GEOTCPTCP

02000400060008000

100001200014000160001800020000

0 5000 10000 15000 20000 25000 30000

Qu

eue L

eng

th


Five TCP : SW1 Queue Length

550 ms × 100 Mbps = 130 kCellsQ << Previous loop delay


23

ERICA+ VS/VD GEOERICA+ VS/VD GEOTCPTCP

0

100

200

300

400

500

600

0 5000 10000 15000 20000 25000 30000

Qu

eue L

eng

th


Five TCP : SW2 Queue Length

10 ms × 100 Mbps = 2.4 kCellsQ << Previous loop delay


24

VS/VD Sw. Feedbackdelay

B/wdiff.

MaxExp. Q(cells)

MaxObs. Q(cells)

OFF Sw1 120 0 0 0OFF Sw2 120 100 3*28 K 30 KON Sw1 100 100 3*25 K 30 KON Sw2 10 100 3*2.4 K 3 KOFF Sw1 570 0 0 0OFF Sw2 570 100 3*135 K 140 KON Sw1 550 100 3*130 K 140 KON Sw2 10 100 3*2.4 K 3 K

Simulation ResultsSimulation Results


25

ObservationsObservationsq Without VS/VD:

m Single control loop for the entire connection.

m All queues are in the bottleneck switch.

m Buffer requirements for terrestrial switch areproportional to satellite propagation delay.

q With VS/VD:

m Control loop broken at each switch.

m Queues remain at the switch between the satelliteand the terrestrial loop (satellite switch).

m Terrestrial switch only requires small buffers.


26

SummarySummary

q VS/VD switch architecture:

m Per-VC queues drain at an ACR based only on theexternal congestion and class Q

m Feedback to upstream queue must include externalcongestion, class Q, and per-VC Q.

m Each queue must monitor its input and output rate.


27

Summary (Cont)Summary (Cont)

q With correct implementation of VS/VD:Maximum queue at each switch< Bandwidth delay product of the previous loop⇒ Can help isolate long-delay hops from short-delayhops.

q Workgroup switches on satellite paths will not needbuffering proportional to round-trip even if they arethe bottleneck.

q Motion: Add sample VS/VD scheme to baseline text

Switch 1

Switch 2

S5


28

Future WorkFuture Workq More complex configurations.

q Presence of VBR background.

q Analysis of complexity of VS/VD switch.

q Scheduling policies for per-VC and class queues.


29

MotionMotionq Add the following two paragraphs to

I.5.4 of the baseline text.I.5.4 A Sample Explicit Rate VS/VD Switch Algorithm

One simple method to implement VS/VD is to have a separate queue(per-VC queue) for each VC. A server at the head of each of thesequeues monitors the input rate of the queue, provides feedback to theupstream queue, and controls the output rate of the queue based on thefeedback from the corresponding downstream server. When providingfeedback, each server only allocates up to the rate at which it isallowed to output (ACR). However, if queues are large, the server mayallocate only a part of its ACR to the previous hop so that its queuescan drain quickly. The main features and options of the algorithm aresimilar to the ERICA+ algorithm. ERICA+ is an extension of theERICA algorithm, and uses queue length to dynamically set the targetABR capacity.


30

Motion (Motion (contdcontd.).)

The basic rate allocation algorithm consists of the following steps atthe end of every averaging interval. The port overload is calculated asthe ratio of the total measured service rate of the per-VC queues andthe target ABR capacity. The fair share term for VCs is calculated asthe ratio of the target ABR capacity to the number of active ABRVC. VCshare is calculated for each VC as the ratio of its measuredservice rate to the overload. The ER for each VC is calculated as ER= Min(Max(Fair Share, VC share), ER from downstream node). TheACR at which the VC's queue drains is determined from this ER aswell as the source-end-system rules for the VS. The feedback to theprevious hop for the VC is calculated as a fraction (based on the VC'squeue length) of the calculated ACR.

Documents

97-1086R1: Per-VC Rate Allocation Techniques For ABR ...jain/atmf/ftp/af_vsv2.pdf · The Ohio State University Raj Jain 1 97-1086R1: Per-VC Rate Allocation Techniques For ABR Feedback