Queueing and Scheduling Traffic is moved by connecting end-systems to switches, and switches to each other Traffic is moved by connecting end-systems to switches, and switches to each other Switches: packet switches v/s circuit switches and connectionless v/s connection oriented Switches: packet switches v/s circuit switches and connectionless v/s connection oriented Data arriving to an input port of a switch have to be moved to one or more of the output ports Data arriving to an input port of a switch have to be moved to one or more of the output ports

Blocking in Packet Switches Can have both internal and output blocking Can have both internal and output blocking  Internal: no path to output  Output: link unavailable Unlike a circuit switch, cannot predict if packets will block Unlike a circuit switch, cannot predict if packets will block If packet is blocked, must either buffer or drop it If packet is blocked, must either buffer or drop it Dealing with blocking: Match input rate to service rate Dealing with blocking: Match input rate to service rate  Overprovisioning: internal links much faster than inputs  Buffering:  input port  in the fabric  output port  shared memory  hybrid

Input Buffering (input queueing) No speedup in buffers or trunks (unlike output queued switch) No speedup in buffers or trunks (unlike output queued switch) Needs arbiter Needs arbiter Problem: head of line blocking Problem: head of line blocking

Dealing with HOL blocking Per-output queues at inputs Per-output queues at inputs Arbiter must choose one of the input ports for each output port Arbiter must choose one of the input ports for each output port Parallel Iterated Matching Parallel Iterated Matching  inputs tell arbiter which outputs they are interested in  output selects one of the inputs  some inputs may get more than one grant, others may get none  if >1 grant, input picks one at random, and tells output  losing inputs and outputs try again Used in DEC Autonet 2 switch Used in DEC Autonet 2 switch

Output Queueing Don’t suffer from head-of-line blocking Don’t suffer from head-of-line blocking Output buffers need to run much faster than trunk speed Output buffers need to run much faster than trunk speed Can reduce some of the cost by using the knockout principle Can reduce some of the cost by using the knockout principle  unlikely that all N inputs will have packets for the same output Most commonly used mechanism in routers Most commonly used mechanism in routers

Shared Memory Route only the header to output port Route only the header to output port Bottleneck is time taken to read and write multiported memory Bottleneck is time taken to read and write multiported memory Doesn’t scale to large switches Doesn’t scale to large switches

Scheduling Scheduling disciplines: Scheduling disciplines:  resolve contention  allocate  bandwidth  delay  loss  determine the fairness of the network  give different qualities of service and performance guarantees Components: Components:  decides service order  manages queue of service requests Example: consider queries awaiting web server Example: consider queries awaiting web server  scheduling discipline decides service order  and also if some query should be ignored

Scheduling Use scheduling: Use scheduling:  Wherever contention may occur  Usually studied at network layer, at output queues of switches Application types: Application types:  best-effort (adaptive, non-real time)  e.g. , some types of file transfer  guaranteed service (non-adaptive, real time)  e.g. packet voice, interactive video, stock quotes Requirements Requirements  implementation ease: few instructions or hardware  has to make a decision once every few microseconds!  work per packet should scale less than linearly with number of active connections  fairness: protection against traffic hogs  performance bounds: on bandwidth, delay and loss  admission control: needed to provide QoS

Max-Min Fairness Scheduling discipline allocates a resource Scheduling discipline allocates a resource An allocation is fair if it satisfies max-min fairness An allocation is fair if it satisfies max-min fairness Intuitively Intuitively  each connection gets no more than what it wants  the excess, if any, is equally shared AB C AB C Transfer half of excess Unsatisfied demand

Scheduling Design Choices Priority: Priority:  packet is served from a given priority level only if no packets exist at higher levels  highest level gets lowest delay (starvation) Work Conservation: Work Conservation:  conservation law: Σρ i q i = constant; where ρ i = λ i x i ; λ i is traffic arrival rate, x i is mean service time for packet; q i is mean waiting time at the scheduler, for connection i;  sum of mean queueing delays received by a set of multiplexed connections, weighted by their share of the link, is independent of the scheduling discipline  work conserving v/s non work conserving disciplines Service Algorithm: Service Algorithm:  FCFS: bandwidth hogs win; no delay guarantees  service tags: arbitrarily reorder queue; provide guarantees; expensive sorting

Work Conserving v/s Non-Work-Conserving Work conserving discipline is never idle when packets await service Work conserving discipline is never idle when packets await service Non work conserving discipline may be idle even when packets await service Non work conserving discipline may be idle even when packets await service  main idea: delay packet till eligible  Reduces delay-jitter => fewer buffers in network  Choosing eligibility time:  rate-jitter regulator: bounds maximum outgoing rate  delay-jitter regulator: compensates for variable delay at previous hop  Always punishes a misbehaving source  Increases mean delay; Wastes bandwidth; Implementation cost

Scheduling Best-effort Connections Main requirement is fairness Main requirement is fairness Achievable using Generalized Processor Sharing (GPS) Achievable using Generalized Processor Sharing (GPS)  Visit each non-empty queue in turn  Serve infinitesimal from each  GPS is not implementable; we can serve only packets  No packet discipline can be as fair as GPS

Weighted Round Robin Serve a packet from each non-empty queue in turn Serve a packet from each non-empty queue in turn Unfair if packets are of different length or weights are not equal Unfair if packets are of different length or weights are not equal Different weights, fixed packet size Different weights, fixed packet size  serve more than one packet per visit, after normalizing to obtain integer weights Different weights, variable size packets Different weights, variable size packets  normalize weights by mean packet size  e.g. weights {0.5, 0.75, 1.0}, mean packet sizes {50, 500, 1500}  normalize weights: {0.5/50, 0.75/500, 1.0/1500} = { 0.01, , }, normalize again {60, 9, 4}  problem: need to know mean packet size in advance Used in some ATM switches Used in some ATM switches

Weighted Fair Queueing (WFQ) Deals better with variable size packets and weights Deals better with variable size packets and weights Also known as packet-by-packet GPS (PGPS) Also known as packet-by-packet GPS (PGPS) Find finish time of a packet, had we been doing GPS; serve packets in order of their finish times Find finish time of a packet, had we been doing GPS; serve packets in order of their finish times WFQ details: WFQ details:  Suppose, in each round, the server served one bit from each active connection  Round number is the number of rounds already completed  can be fractional  If a packet of length p arrives to an empty queue when the round number is R, it will complete service when the round number is R + p => finish number is R + p  independent of the number of other connections!  If a packet arrives to a non-empty queue, and the previous packet has a finish number of f, then the packet’s finish number is f+p  Serve packets in order of finish numbers Finish time of a packet is not the same as the finish number Finish time of a packet is not the same as the finish number

WFQ continued A queue may need to be considered non-empty even if it has no packets in it A queue may need to be considered non-empty even if it has no packets in it  e.g. packets of length 1 from connections A and B, on a link of speed 1 bit/sec  at time 1, packet from A served, round number = 0.5  A has no packets in its queue, yet should be considered non- empty, because a packet arriving to it at time 1 should have finish number 1+ p A connection is active if the last packet served from it, or in its queue, has a finish number greater than the current round number A connection is active if the last packet served from it, or in its queue, has a finish number greater than the current round number Assuming we know the current round number R, finish number of packet of length p Assuming we know the current round number R, finish number of packet of length p  if arriving to active connection = previous finish number + p  if arriving to an inactive connection = R + p

WFQ: computing the round number Naively: round number = number of rounds of service completed so far Naively: round number = number of rounds of service completed so far  what if a server has not served all connections in a round?  what if new conversations join in halfway through a round? Redefine round number as a real-valued variable that increases at a rate inversely proportional to the number of currently active connections Redefine round number as a real-valued variable that increases at a rate inversely proportional to the number of currently active connections With this change, WFQ emulates GPS instead of bit-by-bit RR With this change, WFQ emulates GPS instead of bit-by-bit RR Iterated deletion: Iterated deletion:  A sever recomputes round number on each packet arrival  At any recomputation, the number of conversations can go up at most by one, but can go down to zero, leading to change in round number  Soln: use previous count to compute round number; if this makes some conversation inactive, recompute; repeat until no conversations become inactive

WFQ Implementation On packet arrival: On packet arrival:  use source + destination address (or VCI) to classify it and look up finish number of last packet served (or waiting to be served)  recompute round number  compute finish number  insert in priority queue sorted by finish numbers  if no space, drop the packet with largest finish number On service completion On service completion  select the packet with the lowest finish number Pros: like GPS, provides fairness and protection; can obtain worst-case end-to-end delay bound Pros: like GPS, provides fairness and protection; can obtain worst-case end-to-end delay bound Cons: needs per-connection state; requires a priority queue Cons: needs per-connection state; requires a priority queue Used in most CISCO routers Used in most CISCO routers

Scheduling Guaranteed-Service Connections With best-effort connections, goal is fairness With best-effort connections, goal is fairness Guaranteed-service scheduling Guaranteed-service scheduling  WFQ: provides performance (end-to-end delay) guarantees  Delay-Earliest Due Date:  Earliest-due-date: packet with earliest deadline selected  Delay-EDD prescribes how to assign deadlines to packets  A source is required to send slower than its peak rate  Bandwidth at scheduler reserved at peak rate  Deadline = expected arrival time + delay bound  Delay bound is independent of bandwidth requirement  Implementation requires per-connection state and a priority queue  Rate-controlled scheduling: Regulator shapes the traffic, scheduler provides performance guarantees

Packet Dropping Packets that cannot be served immediately are buffered Packets that cannot be served immediately are buffered Full buffers => packet drop strategy Full buffers => packet drop strategy Packet losses happen from best-effort connections Packet losses happen from best-effort connections Shouldn’t drop packets unless imperative (wasted resources) Shouldn’t drop packets unless imperative (wasted resources) Common strategies: Common strategies:  aggregation: classify packets into classes and drop packet from class with longest queue  priorities: drop lower priority packets  endpoint or regulator marks CLP bit in packets  if network has capacity, all traffic is carried, else dropped  separating priorities within a single connection is hard  early drop: drop even if space is available  drop position: drop packet from some position in the queue

Early Random Drop and RED Early drop => drop even if space is available Early drop => drop even if space is available  signals endpoints to reduce rate  cooperative sources get lower overall delays, uncooperative sources get severe packet loss Early random drop Early random drop  drop arriving packet with fixed drop probability if queue length exceeds threshold  intuition: misbehaving sources more likely to send packets and see packet losses Random early detection (RED) makes three improvements: Random early detection (RED) makes three improvements:  Metric is moving average of queue lengths  Packet drop probability is a function of mean queue length  Can mark packets instead of dropping them  RED improves performance of a network of cooperating TCP sources  small bursts pass through unharmed  prevents severe reaction to mild overload  allows sources to detect network state without losses  controls queue length regardless of endpoint cooperation

Drop Position Can drop a packet from head, tail, or random position in the queue Can drop a packet from head, tail, or random position in the queue Tail: easy; default approach Tail: easy; default approach Head: harder; lets source detect loss earlier Head: harder; lets source detect loss earlier Random: hardest; if no aggregation, hurts hogs most Random: hardest; if no aggregation, hurts hogs most