Flow Aware Networking

Flow Aware Networking

Dominik Najder © 2007 Katedra Telekomunikacji AGH


Queue Model


© 2007 Katedra Telekomunikacji AGHDominik Najder slide 2/18

Table of contents:

Traffic conditioning mechanisms

Scheduling mechanisms

Is it possible to implement DiffServ congestion solutions into FAN?

Fan Router

Priority Fair Queuing

Priority Deficit Round Robin

PFQ vs PDRR

What is the future?



Traffic conditioning mechanisms

4 types of operation

The last mechanisms before scheduling

Traffic conditioning mechanisms| Scheduling mechanisms| DiffServ congestion solutions in FAN| Fan Router| PFQ| PDRR| PFQ vs PDRR| What is the future?



Scheduling mechanisms

Assures differential quality in routing packets assigned to different flows

Different packet types Different quality of service

Many algorithms (FCFS, PS, FQ etc.)

Quality

Selecting which packets should be sent first Selecting which packets may be dropped




Is it possible to implement DiffServ congestion solutions into FAN?

Differences in congestion control ideas:

DiffServ FAN

AC, classifier, traffic conditioner, scheduler blocks Only AC and scheduler blocks – cross protect router

Admission control performed only in edge routers Admission control performed in every router

No information about flows inside router memory – QoS based on information about agregates

Router holds only flow list, without declaration info for traffic class

Domain agent Bandwidth Broker No central agents

No info about network state hold in routers Information based on measurements held in every router




Measurement Based Admission Control (MBAC)




Cross-protect in FAN router

Incoming packets Outgoing packets




Priority Fair Queuing

Modified version of Start-time Fair Queuing

Push-In First-Out queue, each element has timestamp

Active Flow list

Virtual Time counter




SFQ algorithm

Packet arrival:Packet arrival:

1. on arrival of l-byte packet p of flow f:

2. if f is a registered in ActiveList do

3. p.TimeStamp = f.FinishTag

4. f.FinishTag += l

5. else

6. add f to ActiveList

7. p.TimeStamp = VirtualTime

8. f.FinishTag = VirtualTime + l

Packet departure:Packet departure:

1. transmit packets in increasing TimeStamp order

2. at the start of transmission of packet p:

3. VirtualTime = p.TimeStamp

4. for all flows f registered in ActiveList

5. if (f.FinishTag < VirtualTime)

6. remove f

7. if no packets in scheduler VirtualTime = 0




PFQ algorithm

Packet arrival:Packet arrival:1. if PIFO congested, reject packet p at head of longest backlog

2. if f is registered in flow list F

3. if bytes >= MTU

4. p.TimeStamp = f.FinishTag

5. else begin

6. p.TimeStamp = virtual time

7. put behind P, update P

8. f.FinishTag += l

9. else

10. p.TimeStamp = VirtualTime

11. put behind P, update P

12. if flow list is not saturated

13. add flow

14. f.FinishTag = virtual time + l

Main disadvantage:Main disadvantage:

SCALABILTYSCALABILTY

Per packet computational Per packet computational complexity is complexity is O (logQ), where:O (logQ), where:

Q – number of flows server by the Q – number of flows server by the routerrouter




Priority Deficit Round Robin

Based on standard DRR

Per flow f state:

– f.Identier - the flow identifier (possibly a hash of the relevant header fields)

– f.Queue - current length in bytes of flow f queue

– f.Quantum - value of flow f quantum ( >= MTU bytes)

– f.Deficit - current flow deficit

– f.FIFO - addresses of head and tail packets of a linked list forming the flow f FIFO

– f.Next - the next flow in the DRR schedule following flow f




Priority Deficit Round Robin simplified algorithm

Packet Packet enqueueenqueue::1. If no free buffers left then

2. dropPacket();

3. i = ExtractFlow(p);

4. If (i is not registered in AFL)

5. InsertActiveList(i);

6. i.DC = 0;

7. i.ByteCount = size(p);

8. Enqueue(PQ, p)

9. Else

10. i.ByteCount += size(p);

11. If (i.ByteCount <= i.Q) then

12. Enqueue(PQ, p);

13. Else

14.Enqueue(i.Queue, p);

Packet Packet dequeuedequeue::1. While (PQ not empty) do

2. p = Dequeue(PQ); i = ExtractFlow(p);

3. Send(p); i.DC -= Size(p);

4. If (AFL is not empty) then

5. Get head of AFL - flow i;

6. i.DC += i.Q;

7. While ((i.DC >= 0) and (Queue i not empty)) do

8. PacketSize = Size (Head(i.Queue));

9. If (PacketSize <= i.DC) then

10.Send(Dequeue(i.Queue));

11. i.DC -= PacketSize;

12. Else

13.break;

14. RemoveActiveList(i);

15. If (Queue i is not empty) then

16. InsertActiveList(i);




Deficit Round Robin simple example introduction

1200

800

1500

1000

Traffic

Source 1

Traffic

Source 2

Traffic

Source 3

Quantum




Deficit Round Robin simple example, round 1

1200

800

1500

1000

Traffic

Source 1

Traffic

Source 2

Traffic

Source 3

Quantum

Deficit Packet served

1000 No

1000 – 800 = 200 Yes

1000 No




Deficit Round Robin simple example, round 2

1200

800

1500

1000

Traffic

Source 1

Traffic

Source 2

Traffic

Source 3

Quantum



1000 No 2000-1200 = 800 Yes

1000 – 800 = 200 Yes 0 No

1000 No2000–1500 = 500

Yes




PFQ vs PDRR

Which solution is better?

PFQ:PFQ:

-simple

-but higher computational complexity O(logQ) – not well scalable

PDRR:PDRR:

-more complex, bigger amount of data to be held inside router

-computation complexity O(1) – very scalable, main FAN feature over DiffServ




What is the future?

Redesign of IPv6 packet

Minimized header overhead and reduced header process for the majority of packets

Less essential fields removed or moved to extension headers

Traffic class in IPv6= TOS in IPv4

New Flow Label




Thank you

Documents

Flow Aware Networking