Network Model for Delay-Sensitive Traffic

Transcription

1 Traffic Scheduling

2 Network Model for Delay-Sensitive Traffic Source Switch Switch Destination Flow Shaper Policer (optional) Scheduler + optional shaper Policer (optional) Scheduler + optional shaper cfla. Varma 4 1

3 Traffic Shaping Goal: Smooth out bursty traffic to prevent buffer overflows in the network. Also enables network to provide end-to-end delay guarantees Several approaches Leaky bucket Token bucket Token bucket with spacer (dual leaky bucket) cfla. Varma 4 2

4 Simple Leaky Bucket Rate Control cells Leaky bucket capacity σ cells transmitted at rate ρ Enforces minimum spacing between cells Algorithm: Arriving cells queued in leaky bucket Cells transmitted from bucket at rate ρ. Arriving cell discarded when bucket full. Used to shape CBR traffic Easy to police: Need to verify only spacing between cells cfla. Varma 4 3

5 Token Bucket Rate Control rate ρ Token bucket capacity σ cells get token Peak rate of traffic C Tokens generated at fixed rate ρ and placed into token bucket of capacity ff. When bucket full, new tokens are discarded. Cell must obtain token from token bucket before transmission. Limits number of bits transmitted in T interval (ff + T to ρ). cfla. Varma 4 4

6 Token Bucket with Spacer rate ρ Token bucket capacity σ cells get token Leaky bucket capacity c cells transmitted at rate q Combination of leaky bucket and token bucket Limits the instantaneous transmission rate to service rate from leaky bucket. Used to shape VBR traffic cfla. Varma 4 5

7 Generic Cell-Rate Algorithm (GCRA) Algorithm defined by ATM Form Traffic Management Spec for policing CBR and VBR traffic Arrival of cell k from connection at time t(k) TAT = Theoretical Arrival Time L = Limit I = Increment TAT < t(k)? Y N TAT = t(k) Non conforming cell Y TAT > t(k) + L? N TAT = TAT + I Conforming cell cfla. Varma 4 6

8 Traffic Scheduling 1 2 n Problem: Schedule transmission of packets on an outgoing link to meet QoS guarantees. Packets may be fixed or variable length Difficulties: Multiple constraints (delay, bandwidth, jitter, loss rate) Complexity of implementation cfla. Varma 4 7

9 Applications of Scheduling Algorithms Useful in many environments: Providing isolation to traffic requiring guarantees (example: Real-time video session requiring delay guarantee) Distributing available bandwidth to sessions fairly Implementing hierarchical link sharing cfla. Varma 4 8

10 Desirable Attributes of a Scheduling Algorithm Low deterministic delay bound Affects output burstiness and buffer requirements Isolation from behavior of other sessions Delay bound independent of number of sessions Ability to efficiently control delay bound Fairness Distribution of free bandwidth Worst-case fairness Accuracy of algorithm in small timescales Complexity of implementation cfla. Varma 4 9

11 Classification of Traffic Schedulers Based on scheduling discipline: Work-conserving: Server never remains idle when a packet is in system. Non-work-conserving: Packets may wait even if server is idle. Based on implementation: Sorted priority: Timestamp associated with each packet. ffi Packets transmitted in the order of their time-stamps. ffi Frame based: Time split into frames. ffi Upper bound on number of packets from each connection ffi serviced during frame period. cfla. Varma 4 10

12 Scheduler Classification (continued) Based on type of guarantee 1. Algorithms based on rate guarantee Guarantees minimum service rate to a session Explicit bandwidth ) guarantee implicit delay guarantee Delay a function of allocated rate 2. Algorithms based on guaranteed rate and delay Allows different maximum delays to be guaranteed for the same rate Low-bandwidth sessions can still obtain low delays cfla. Varma 4 11

13 Some Proposed Scheduling Schemes Frame-based schedulers Round-robin (Weighted round-robin, deficit round-robin, hierarchical round-robin, etc.) Stop-and-Go queueing Sorted-priority (timestamp-based) schedulers Generalized Processor Sharing (GPS) Fair queueing and weighted fair queueing VirtualClock Self-clocked fair queueing Frame-based Fair Queueing Many other fair queueing algorithms cfla. Varma 4 12

14 Weighted Round-Robin Scheduler VC 0 50 Weighted Round Robin Scheduler Virtual Channels VC 1 1 Packet Transmission VC 50 1 Optimal when the sessions are allocated equal rates, but Limited flexibility in allocating bandwidth Worst-case delay proportional to size of round cfla. Varma 4 13

15 An Example Sorted-Priority Scheduler Timestamp calculated as departure time of last bit of packet in fluid model. Timestamp of current packet of i session = Timestamp of previous packet of i session Length of current packet Allocated rate of i session + Packet transmissions in order of timestamp values Session 1 50 TS=8 TS=6 TS=4 TS=2 Session 2 25 TS=12 TS=8 TS=4 Session 3 25 TS=12 TS=8 TS=4 Key property: Packet with timestamp = t would have departed by time t (assuming all sessions remain backlogged). cfla. Varma 4 14

16 An Example Sorted-Priority Scheduler (continued) Session 1 Session 2 Outgoing traffic Time 2000 Problem: Sessions using left-over bandwidth will be punished ) later Poor fairness Timestamp calculations must take into account global state for sessions to use excess bandwidth without future starvation. cfla. Varma 4 15

17 Generalized Processor Sharing (GPS) Flow 1 Flow 2 Output link Flow 3 Used as the baseline to compare other algorithms Assumes packets are infinitely divisible (fluid model) Flow i allocated a guaranteed rate of g i Backlogged connections serviced at each instant of time with rate proportional to their reservations Weighted Fair Queueing: Packet-level approximation of GPS Finishing times of packets under GPS used to order transmissions cfla. Varma 4 16

18 Tracking Global State in a WFQ Scheduler Virtual-time function v(t) maintains global state Piecewise linear function of real t time Slope depends on allocated rates of backlogged sessions. V Assume sessions, and allocated rate of i session ρ = i. Slope v(t) of during the (t1;t2) interval given by VX X i2b(t 1 ;t 2 ) ρ i; where B(t1;t2) set of backlogged sessions during the interval. j=1 ρ j= cfla. Varma 4 17

19 Tracking Global State in a WFQ Scheduler: Example Three sessions with bandwidth allocations 50, 25, and 25 percent. Link rate = 100,000 cells/second. cfla. Varma 4 18

20 Tracking Global State in a WFQ Scheduler: Example Backlog of session 1 Backlog of session 2 Backlog of session 3 slope = 2 slope = 4 virtual time v(t) slope = 1 slope = 1.33 slope = time t (microseconds) cfla. Varma 4 19

21 k ψ max(ts k 1 i ; Real Time) + L TS : i i ρ VirtualClock Attempts to simulate time-division multiplexing. On arrival, packet stamped with time-stamp based on its arrival time and average rate. Packet time-stamps increase at allocated rate. Time-stamp of idle connection initialized from real time. Difference of time-stamp from real time unbounded Transmissions scheduled based on time-stamp. cfla. Varma 4 20

22 Self-Clocked Fair Queueing Proposed by Golestani to simplify time-stamp computation in Weighted Fair Queueing. Basic idea: Use time-stamp of packet currently under transmission to compute time-stamp on incoming packet Simple to implement, but delay bound grows with number of connections. cfla. Varma 4 21

23 End-to-end Delay Bound in a Fluid Server Switch 1 Switch 2 Switch k Source of session i σ i bits capacity = ρ i bits/sec Max delay in ideal fluid system = ff i =ρ i. cfla. Varma 4 22

24 End-to-end Delay Bound in a WFQ Server Entire packet must be received before forwarding ) Delay of i =ρ i per switch. L Packet server can be behind fluid server by L max =r, where L max is the maximum size of a packet across all sessions. Max delay in a network of k servers: D» (ff i =ρ i ) + (k 1)(L i =ρ i ) + k(l max =r) cfla. Varma 4 23

25 References H. Zhang, Service disciplines for guaranteed performance service in packet-switching networks, Proceedings of the IEEE, vol. 83, pp , October A.Demers, S. Keshav, and S.Shenker, Analysis and simulation of a fair queueing algorithm, Journal of Internetworking Research and Experience, vol. 1, no. 1, pp. 3 26, September D. Stiliadis, A. Varma, Rate-Proportional Servers: A General Methodology for the Design of Fair Queueing Algorithms, IEEE/ACM Transactions on Networking, April D. Stiliadis, A. Varma, Efficient Fair-Queueing Algorithms for ATM and Packet Networks, IEEE/ACM Transactions on Networking, April S. J. Golestani, A self-clocked fair queueing scheme for broadband applications, Proceedings of IEEE INFOCOM 94, pp D. Stiliadis, A. Varma, Latency-Rate Servers: A General Model for Analysis of Traffic Scheduling Algorithms, IEEE/ACM Transactions on Networking, October M. Shreedhar and G. Varghese, Efficient fair queueing using deficit round robin, in Proc. ACM SIGCOMM 95, pp , September cfla. Varma 4 24