QoS provisioning. Lectured by Alexander Pyattaev. Department of Communications Engineering Tampere University of Technology

Transcription

1 QoS provisioning Lectured by Alexander Pyattaev Department of Communications Engineering Tampere University of Technology March 6, 2012

2 Outline 1 Introduction 2 QoS support elements Traffic classification Traffic monitoring Traffic shaping Connection admission control Scheduling for service Queue management Backup and reservation Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

3 Provisioning QoS? Provisioning = providing resources in advance So we provide resources that QoS-aware system will use to provide services with guaranteed quality level. General law is - we provide more then will actually be used This is called overprovisioning This is better than running out of resources! Our task is to compute exactly the required amount And on top of that the minimal required reserve for possible failures Secondary tasks include: Fair/priority scheduling - decide who goes first Admission control - decide who is VIP Traffic shaping/network feedback - control the flows at source Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

4 QoS support elements We build a QoS-aware system based on several key elements: 1 Traffic classification 2 Traffic monitoring 3 Traffic shaping 4 Connection admission control 5 Scheduling for service 6 Queue management 7 Backup and reservation We also build it in such a way that we have a clear performance limit, which we would then advertise to the customers. The goal of this lecture is to provide an overview of what is available to us, before we dive into the complex topics behind. Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

5 Outline 1 Introduction 2 QoS support elements Traffic classification Traffic monitoring Traffic shaping Connection admission control Scheduling for service Queue management Backup and reservation Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

6 Traffic classes and classification Typically, networks carry more then one class of traffic, e.g. Voice data - CBR class Video broadcasts - nrt class Bulk transfers - BE class Separate QoS agreements should be made for each class, and for all of them together Sources Monitoring CBR CBR Priority Uplink service BE Classifier BE nrt nrt Drop Normal service Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

7 Classification approaches One may approach classification from different angles: Agree during connection setup (RSVP protocol does that) Include class of servie label into every packet (IP) Classify traffic dynamically (How exactly is a separate question) First two approaches are widely accepted and are used widely. The dynamic classification typically is employed when some prioritization is needed, for example: Connections slower than 100 kbps - high priority (nrt service) 100 kbps - 1Mbps - best-effort service More than 1Mbps - drop packets Such policies are typically employed when no prior knowlege of the traffic can be expected Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

8 Traffic monitoring Monitoring means measurements. Primary application is to provide data for policing, shaping and queue management. The classification can be based on measurements (as described above). We have to measure: Sustained rate Bursts and their parameters Delay and jitter This has to be done on a per-flow basis, so it has to be Fast and scalable Stable and autonomous Accurate Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

9 Token bucket A universal all-in-one tool for measurements of rate Deep inside is in fact a primitive queuing system model + some extra logic Parameters: B - bucket size R - arrival rate S - service rate T 1 - almost-full threshold T 2 - almost-empty threshold See picture on the next slide Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

10 Token bucket T 1 Arrivals R tokens/s B T 2 c(t) Departures S tokens/s Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

11 Using token buckets - monitor flow rate This is the most common usecase. 1 Fix the token arrival rate R, tokens arrive at regular intervals 2 Every time the packet of size s arrives on a link, remove s tokens If c(t) < s mark the packet as non-conforming 3 Repeat from step 1 Non-conforming packets are subjected to the appropriate policy later. Bucket size B controls the system inertia: If B < s min, all packets will be marked If B s, s const, only CBR flows will pass through If B = n s max, bursts of up to n packets will be allowed to pass One may use analytical results for M/D/1 system to analyze the performance of this algorithms in first approximation, yet G/D/1 model suits much better. Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

12 Using token buckets - marking fluctuating traffic Single Rate Three Color Marker (srtcm), RFC 2697 Traffic parameters that are controlled: Committed Information Rate, CIR Committed (expected) Burst Size, CBS Excess Burst Size, EBS srtcm classifies to one of three categories Green: traffic does not exceed (CIR, CBS) Yellow: exceeds (CIR,CBS), but not (CIR,EBS) Red: exceeds (CIR,EBS) Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

13 Details of srtcm CIR - measured in bytes/s. or packets/s - expected maximum sustained rate CBS and EBS - measured in bytes or packets - expected and maximum burst parameters, some restrictions apply: (EBD > 0) (CBS > 0) and EBS CBS If bytes are used, max(cbs, EBS) s max Basis of srtcm: two Token Bucket elements, connected in chain. Arrivals TB(CIR,EBS) TB(CIR,CBS) Classified traffic Overflow Overflow Here TB(B, R) is a token bucket with size B and rate R Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

14 Three-rate TCM (trtcm) Defined in RFC Like srtcm, but also controls peak information rate (PIR). trtcm classifies to the same three categories: Green: traffic does not exceed (CIR, CBS) Yellow: exceeds (CIR,CBS), but not (PIR,EBS) Red: exceeds (PIR,EBS) Arrivals TB(CIR,EBS) TB(PIR,CBS) Classified traffic Overflow Overflow TCM algorithms can be extended to fit more rates by putting more token buckets in chain, which allows to construct arbitrary classifiers. Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

15 Measuring delay In most cases, delay is explicitly measured by sending special probe packets into the network. Refer to RFC 2544 for some examples for Ethernet networks. There are 2 basic approaches: Two devices with synchronized clocks (typically via GPS) are put at the ends of the network, time delta is measured directly. This typically requires specialized hardware. A loopback device is put at one end, a round-trip-time(rtt) is measured. This is what ping utility does, and requires no extra hardware. In both cases measurements have to be performed on a loaded network. clock source loopback Network Network Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

16 What about delay jitter? Unlike delay itself, delay jitter can be easily measured. Delay jitter is essentially the derivative of inter-arrival times. Again, refer to RFC 2544 for some examples for Ethernet networks. This is the main reason why it is used for QoS instead of variance Normally, a mean square value over some window is reported An absolute (or squared) value of delay jitter may be used to feed token bucket In this case single delayed packets would not trigger an alarm Yet a consistent delay disturbances would! Jitter can be also measured along with delay when 2 devices with synchronized clocks are used. More stuff to come in the lecture on traffic monitoring and policing! Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

17 Traffic shaping concept The shaping implies changing the traffic pattern without changing the total quantity. Shaping is implemented through smart buffering, that is passing the packets through a delay line. Shaping allows to remove bursts and load spikes This comes at a cost of extra delay! Shaping is equivalent to lowpass-filtering of arrival flow rate Traffic shaping is mostly employed in core networks, where multiple flows are aggregated and reshaped to fit a given backbone. There will be a separate lecture on buffering and buffer management! Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

18 Traffic shaping for a broadband link 5 4 Original flow Shaped flow Possible losses Link capacity Here you can see how link usage can be optimized with traffic shaping Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

19 Time-sequence chart for shaping Cumulative flow, pkts Traffic shaping - time-seq chart Arriving flow Shaped flow Possible losses This proves that no traffic is discarded, it is merely delayed (for 5 sec) Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

20 Admission control idea Question: is there a way to support this traffic? Motivation: it is better to serve few customers properly, then all of them badly. The motivation is reasonable only for very sensitive services, especially voice calls. Connection admission control (CAC) variants: Dynamic Static Combination of these two CAC can be enforced for: Single flow Aggregated flow Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

21 Dynamic CAC Usually used for single flows CAC function is evaluated at each network node separately Special resource reservation protocols are used (RSVP, etc.) If any router along the path refuses reservation, it is rejected completely Incorporated in Integrated Services, ATM, MPLS, SS7. Dynamic CAC is typically preferred for dynamic networks, but it may be too slow for highly dynamic networks like vehicular mesh networks. Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

22 Static CAC Usually used for aggregated flows CAC function is evaluated at network entry point Static CAC is used between ISP s and telecoms operators Static CAC is also used for voice connections in telephone networks Static CAC is used for static setups like broadband links, or for services that require such behavior. Such services include telephony, video conferences, gaming etc. There will be a separate lecture on admission control! Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

23 Scheduling concept There are always some shared resources. We would like to distribute those among many users: buffer space rate of the outgoing link processing time radio resources in radio networks Task of a scheduling algorithm is to distribute those resources according to some policy between users. There are many ways to do that! Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

24 Scheduler types Schedulers without priorities FIFO Round Robin, RR Deficit Round Robin, DRR Priority schedulers Generalized Processor Sharing, GPS Weighed Fair Queuing, WFQ Self-Clocked Fair Queuing, SCFQ Completely Fair Queuing, CFQ Worst-Case Fair Weighed Fair Queuing, WF2Q Worst-Case Fair Weighed Fair Queuing Plus, WF2Q+...just to name few! Note: There are many more, only few are actually implemented There will be a separate lecture on admission control! Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

25 Queue management Idea: If the queue overflow is imminent, when and which packet should be dropped? AIM: congestion(overflow) control/avoidance Performance of the algorithm is evaluated based on: Congestion control performance Decision whether to drop is taken When packet arrives to the system When congestion is experienced Queue management algorithms Passive Droptail (Taildrop) Random (Just drop random packets) Active Random early detection (RED)...and many more There will be a separate lecture on traffic policing! Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

26 Reservation basics Reserves are the resources that are used when something breaks. Affect the reliability only. Several policies are possible: N + 1 reservation - for each element keep 1 spare N + x reservation - for each element keep x spare ones 2N reservation - keep a complete system duplicate For example, the total reliability for an N+1 system built of blocks with reliability p 0 could be estimated as follows: p ok = 1 ((1 p 0 ) (N + 1) (1 p r )), where p r is the probability that reserves are activated properly. It is critical to realize that 2N reservation, although providing excellent resilience, is typically very expensive, and should be avoided at all costs. One has to compute the required reservation rate for each component to guarantee reliability. Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27

27 Conclusions We have considered building blocks for QoS Those allow us to draft and test a network that provides QoS Those do not tell us how to dimension such network! There are several ways to put the blocks together, those are discussed in the lecture on QoS frameworks. Some of the topics discussed require a more detailed look, and will be discussed separately. Lectured by Alexander Pyattaev (TUT) TLT-2727 March 6, / 27