Congestion Control and Traffic Management in High Speed

Congestion Control and Traffic Management in High Speed Networks Carey Williamson University of Calgary 1

Introduction l The goal of congestion control is to regulate traffic flow in the network in order to avoid saturating or overloading intermediate nodes in the network 2

Congestion: Effects Congestion is undesirable because it can cause: Increased delay, due to queueing within the network Packet loss, due to buffer overflow Reduced throughput, due to packet loss and retransmission l Analogy: “rush hour” traffic l 3

Congestion: Causes The basic cause of congestion is that the input traffic demands exceed the capacity of the network l In typical packet switching networks, this can occur quite easily when: - output links are slower than inputs - multiple traffic sources competing for same output link at the same time l 4

Buffering: A Solution? Buffering in switches can help alleviate short term or transient congestion problems, but. . . l Under sustained overload, buffers will still fill up, and packets will be lost l – only defers the congestion problem l More buffers means more queuing delay – beyond a certain point, more buffering makes the congestion problem worse, because of increased delay and retransmission 5

Motivation The congestion control problem is even more acute in high speed networks l Faster link speeds mean that congestion can happen faster than before e. g. , 64 kilobyte buffer @ 64 kbps: 8. 2 seconds @ 10 Mbps: 52 milliseconds @ 1 Gbps: 0. 52 milliseconds l 6

Motivation (Cont’d) l Buffer requirements increase with link speeds e. g. , to store 1 second worth of traffic @ 64 kbps: 8 kilobytes @ 10 Mbps: 1. 25 Mbytes @ 1 Gbps: 125 Mbytes 7

Motivation (Cont’d) Heterogeneity of link speeds - just because you add new high speed links to a network doesn’t mean that the old low speed links go away - interconnecting high speed and lower speed networks creates congestion problems at the point of interconnect l 8

Motivation (Cont’d) l Traffic is bursty - high peak-to-mean ratio, peak rates - e. g. , data traffic: 10 -to-1, 1 -10 Mbps - e. g. , video traffic: 20 -to-1, 5 -100 Mbps - can statistically multiplex several channels, but if too many are active at the same time, congestion is inevitable 9

Motivation (Cont’d) Reaction time is bounded by the propagation delay - in a high-speed wide-area network, the delay x bandwidth product is HUGE!!! - d x b tells you how many bits fit in the “pipe” between you and the receiver - by the time you realize that network is congested, you may have already sent another Mbit or more of data!!! l 10

Reactive versus Preventive There are two fundamental approaches to congestion control: reactive approaches and preventive approaches l Reactive: feedback-based l – attempt to detect congestion, or the onset of congestion, and take action to resolve the problem before things get worse l Preventive: reservation-based – prevent congestion from ever happening in the first place, by reserving resources 11

Reactive versus Preventive (Cont’d) l Most of the Internet approaches are reactive schemes – TCP Slow Start – Random-Early-Detection (RED) Gateways – Source Quench The large d x b product means that many of these approaches are not applicable to high speed networks l Most ATM congestion control strategies are preventive, reservation-based l 12

Congestion Control in ATM l When people discuss congestion control in the context of high speed ATM networks, they usually distinguish between call-level controls and cell-level controls 13

Call-Level Control An example of the call-level approach to congestion control is call admission control (to be discussed later this semester) l Tries to prevent congestion by not allowing new calls or connections into the network unless the network has sufficient capacity to support them l 14

Call-Level Control (Cont’d) At time of call setup (connection establishment) you request the resources that you need for the duration of the call (e. g. , bandwidth, buffers) l If available, your call proceeds l If not, your call is blocked l E. g. , telephone network, busy signal l 15

Call-Level Control (Cont’d) Tradeoff: aggressive vs conservative l Want to accept enough calls to have reasonably high network utilization, but don’t want to accept so many calls that you have a high probability of network congestion (which might compromise the QOS requirements that you are trying to meet) l 16

Call-Level Control (Cont’d) l Problems: Can be unfair - denial of service, long access delay Hard to specify resource requirements and QOS parameters precisely - may not know, or may cheat - congestion can still occur 17

Cell-Level Control Also called input rate control l Control the input rate of traffic sources to prevent, reduce, or control the level of congestion l Many possible mechanisms: Traffic shaping, traffic policing, UPC Leaky bucket (token bucket) Cell tagging (colouring), cell discarding Cell scheduling disciplines l 18

Congestion Control in ATM There is actually a complete spectrum of traffic control functions, ranging from the very short-term (e. g. , traffic shaping, cell discarding) to the very long-term (e. g. , network provisioning) l See [Gilbert et al 1991] l 19

ATM Traffic Control Schemes Time Scale 20

ATM Traffic Control Schemes Short Term usec Time Scale 21

ATM Traffic Control Schemes Long Term Short Term Months, years usec Time Scale 22

ATM Traffic Control Schemes Long Term Resource Provisioning Call Duration Propagation Delay Time Cell Time Scale Admission Control Routing, Load Balancing Explicit Congestion Notification Fast Reservation Protocol Node to Node Flow Control Usage Parameter Control Priority Control Traffic Shaping Cell Discarding 23

ATM Traffic Control Schemes Cell Time Scale Usage Parameter Control Priority Control Traffic Shaping Cell Discarding 24

ATM Traffic Control Schemes Propagation Delay Time Explicit Congestion Notification Fast Reservation Protocol Node to Node Flow Control Time Scale 25

ATM Traffic Control Schemes Call Duration Admission Control Routing, Load Balancing Time Scale 26

ATM Traffic Control Schemes Long Term Resource Provisioning Time Scale 27

ATM Traffic Control Schemes Long Term Resource Provisioning Call Duration Propagation Delay Time Cell Time Scale Admission Control Routing, Load Balancing Explicit Congestion Notification Fast Reservation Protocol Node to Node Flow Control Usage Parameter Control Priority Control Traffic Shaping Cell Discarding 28

ATM Traffic Control Schemes l Preventive controls: Resource provisioning Connection admission control Call routing and load balancing Usage parameter control Priority control Traffic shaping Fast reservation protocol 29

ATM Traffic Control Schemes l Reactive controls: Adaptive admission control Call routing and load balancing Adaptive usage parameter control Explicit congestion notification (forward or backward) Node to node flow control Selective cell discarding 30

Leaky Bucket One of the cell-level control mechanisms that has been proposed is the leaky bucket (a. k. a. token bucket) l Has been proposed as a traffic policing mechanism for Usage Parameter Control (UPC), to check conformance of a source to its traffic descriptor l Can also be used as a traffic shaper l 31

Leaky Bucket (Cont’d) Think of a bucket (pail) with a small hole in the bottom l You fill the bucket with water l Water drips out the bottom at a nice constant rate: drip, drip. . . l 32

Leaky Bucket (Cont’d) 33

Leaky Bucket (Cont’d) Bucket 34

Leaky Bucket (Cont’d) Empty Bucket 35

Leaky Bucket (Cont’d) Bucket Hole 36

Leaky Bucket (Cont’d) Water Bucket Hole 37

Leaky Bucket (Cont’d) 38

Leaky Bucket (Cont’d) Drip 39

Leaky Bucket (Cont’d) 40

Leaky Bucket (Cont’d) 41

Leaky Bucket (Cont’d) 42

Leaky Bucket (Cont’d) Constant rate stream of drips, all nicely spaced, periodic 43

Leaky Bucket (Cont’d) Storage area for drips waiting to go Constant rate stream of drips, all nicely spaced, periodic 44

Leaky Bucket (Cont’d) A leaky bucket flow control mechanism is then a software realization of this very simple idea l Packets (cells) waiting for transmission arrive according to some (perhaps unknown) arrival distribution l Tokens arrive periodically (deterministically) l Cell must have a token to enter network l 45

Leaky Bucket (Cont’d) Incoming Tokens at rate r tokens/sec Incoming Cells (generated by traffic source with rate X) + 46 To Network

Leaky Bucket (Cont’d) Incoming Tokens at rate r tokens/sec Incoming Cells 5 4 32 + 1 47 To Network

Leaky Bucket (Cont’d) Incoming Tokens Incoming Cells 5 4 32 + 1 48 To Network

Leaky Bucket (Cont’d) Incoming Tokens Incoming Cells 5 + 43 2 49 1 To Network

Leaky Bucket (Cont’d) Incoming Tokens Incoming Cells 5 + 4 3 50 2 1 To Network

Leaky Bucket (Cont’d) Incoming Tokens Incoming Cells 5 + 4 51 3 2 1 To Network

Leaky Bucket (Cont’d) Incoming Tokens Incoming Cells 5 X + 4 52 3 2 1 To Network

Leaky Bucket (Cont’d) Incoming Tokens Incoming Cells XX 5 + 4 53 3 2 1 To Network

Leaky Bucket (Cont’d) Incoming Tokens Incoming Cells XX + 54 3 2 1 To Network

Leaky Bucket (Cont’d) Cells must obtain tokens in order to proceed into the network l If no token available, then cell is discarded l Constrains the rate at which cells enter the network to be the rate negotiated at the time of call setup l Shapes traffic, reduces burstiness l 55

Buffered Leaky Bucket Arriving cells that find a token waiting can proceed directly into the network l Arriving cells that find no token ready must wait in queue for a token l Cells that arrive to a full queue are lost l Tokens that arrive to a full token pool are simply discarded l 56

Buffered Leaky Bucket Incoming Tokens at rate r tokens/sec Pool of at most M waiting tokens Incoming Cells + Queue of at most B waiting cells 57 To Network

Buffered Leaky Bucket (Cont’d) Incoming cell rate: X l Token rate: r l If X > r, then cells wait in buffer until tokens are available Output traffic is r cells/sec, nicely paced l If X < r, then tokens always ready Output traffic is X (< r) l Use for traffic shaping or UPC l 58

Buffered Leaky Bucket (Cont’d) A station can “save up” at most M tokens l Limits the maximum burst size in the network l Can send at most M cells back to back l B can be set to balance the tradeoff between cell loss and cell delay l 59

Leaky Bucket UPC The token rate r is set based on the rate declared at the time of call setup l Makes sure that each source obeys rate that was used when the call admission decision was made (i. e. , descriptor) l Can use “single leaky bucket” to police just the peak cell rate (PCR) l Can use “dual leaky bucket” to police both PCR and SCR l 60

Variations There are several different variations of the basic leaky bucket concept described in the literature, such as the virtual leaky bucket, spacer, others l Basic idea: rather than strictly enforcing rates, allow senders to occasionally exceed their prescribed rate, as long as they mark or tag their extra cells l 61

Cell Marking Scheme Uses leaky bucket to regulate cell transmissions as before, but rather than having cells wait for tokens when there are no tokens ready, the station can transmit the cell and mark it as a violation cell (i. e. , cell colouring) l Green (CLP = 0) for cells that obey rate l Red (CLP = 1) for cells that don’t l 62

Cell Colouring (Cont’d) If the network detects congestion at any point, then it does not hesitate to throw away red cells (CLP = 1), but always gives preference to green cells l Green cells must get through l Red cells get through only if there is spare capacity in the network l “No harm in trying!” principle l 63

Selective Cell Discard (SCD) A cell-level control mechanism in ATM switches called selective cell discard can be implemented quite easily using a CLP threshold on each queue/buffer l Below the threshold, can accept both green and red cells l Beyond the threshold, can only accept green cells l 64

Selective Cell Discard (Cont’d) Buffer in an ATM switch 65

Selective Cell Discard (Cont’d) Buffer in an ATM switch 66

Selective Cell Discard (Cont’d) Some cells waiting to go Buffer in an ATM switch 67

Selective Cell Discard (Cont’d) CLP Threshold Buffer in an ATM switch 68

Selective Cell Discard (Cont’d) CLP Threshold 69

Selective Cell Discard (Cont’d) CLP Threshold 70

Selective Cell Discard (Cont’d) CLP Threshold 71

Selective Cell Discard (Cont’d) CLP Threshold 72

Selective Cell Discard (Cont’d) CLP Threshold 73

Selective Cell Discard (Cont’d) CLP Threshold 74

Selective Cell Discard (Cont’d) CLP Threshold 75

Selective Cell Discard (Cont’d) CLP Threshold 76

Selective Cell Discard (Cont’d) CLP Threshold 77

Selective Cell Discard (Cont’d) CLP Threshold 78

Selective Cell Discard (Cont’d) CLP Threshold 79

Selective Cell Discard (Cont’d) CLP Threshold 80

Selective Cell Discard (Cont’d) CLP Threshold 81

Selective Cell Discard (Cont’d) CLP Threshold 82

Selective Cell Discard (Cont’d) CLP Threshold 83

Selective Cell Discard (Cont’d) CLP Threshold 84

Selective Cell Discard (Cont’d) CLP Threshold 85

Selective Cell Discard (Cont’d) CLP Threshold 86

Selective Cell Discard (Cont’d) CLP Threshold 87

Selective Cell Discard (Cont’d) CLP Threshold 88

Explicit Congestion Notification There are some proposals to use reactive congestion control approaches for end-to-end flow control in ATM l One of the mechanisms proposed is called Explicit Forward Congestion Notification (EFCN) (or EFCI, for Explicit Forward Congestion Indication) l 89

EFCI: Basic Operation Switches can detect the onset of congestion (e. g. , buffers filling up) l Switches set a control bit in cell headers to indicate this congestion condition l Sources react by reducing the volume of traffic that they are sending through that switch l Suitable for VBR or ABR traffic l 90

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch 91 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch Buffer 92 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch Occupied Unoccupied 93

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch EFCI Threshold 94 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch Data Cell EFCI Threshold 95 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch EFCI Threshold 96 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch EFCI Threshold 97 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch EFCI Threshold 98 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch EFCI Threshold 99 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch EFCI Threshold 100 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch EFCI Threshold 101 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch EFCI Threshold 102 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch !!! EFCI Threshold 103

EFCI: Basic Operation (Cont’d) Ack Cell Traffic Source Switch EFCI Threshold 104 Switch Traffic Sink

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch EFCI Threshold 105 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch EFCI Threshold 106 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch EFCI Threshold 107 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch !!! EFCI Threshold 108 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch EFCI Threshold 109 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch EFCI Threshold 110 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch EFCI Threshold 111 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch EFCI Threshold 112 Switch

EFCI: Basic Operation (Cont’d) Traffic Sink Traffic Source Switch EFCI Threshold 113 Switch

EFCI Issues How to set EFCI threshold l What should sources do when EFCI signal is seen l What should sources do when no EFCI signal is seen l Forward versus backward notification l Effect of feedback delay l Delay x bandwidth product l 114

Summary Congestion control in high speed ATM networks is a difficult problem l Lots of good ideas of how to do it, but no real standard (yet? ) l Will likely require a combination of schemes at different time scales and for different classes of traffic l Lots more remains to be done l 115