1	Quality-of-Service: Methods and Software Vincenzo Liberatore Case Western Reserve University
2	Pervasive Computing
3	Real-Time Issues Quality-of-Service (QoS) Ensure low delay, jitter, loss rates, and high bandwidth Implement below, at, or above IP Real-Time Protocol (RTP) Transport protocol Support real-time communication
4	Quality-of-Service Problem Best-effort networks Congestion Metrics Implementation Application-middleware IP Data link
5	Quality-of-Service (QoS) Best-effort model Prevalent in the Internet No end-to-end circuit Packets can be delayed Delay is often unpredictable No bandwidth reservation QoS Ameliorate best-effort model E.g., packet priorities Impact on control Delays E.g., forces small feedback gains Packet drop-outs Infinite delays Jitter Delay variability Often more troublesome than average delays Bandwidth Scalability
6	QoS and Congestion Objectives QoS protects from cross-traffic QoS does not protect from failures, interferences But in some cases, it does Possible strategies Priorities Reservation Weighted Fair Queuing
7	Congestion: Example Scenario One controller Discrete LQR Assumes 50ms sampling rate Assumes no delays n inverted pendulum Bottleneck: T1 line RTT: 3ms Plot response over time Congestion When more plants are added Scalability
8	Congestion and Performance
9	Overprovisioning Over-provisioning Large bandwidth implies short queues 1Gbps Ethernet vs. 1Mbps CAN Useful for probabilistic QoS Risk Congestion caused by even one cross-traffic flow TCP tends to take over Example: ping a machine in same building with and without one FTP flow as cross-traffic Explains poor plant performance with one flow
10	Congestion and TCP One TCP flow can congest a link 1 plant, 1 TCP flow Scalar linear plant Sporadic TCP [R03] Congest link sporadically Average utilization rate low Difficult convergence
11	Overprovisioning Access networks Any flow can saturate the network Exacerbated by TCP Attempt to use all the bandwidth Core network Good news Statistical multiplexing Relatively low rate flows Bad news Not ready yet for VoIP [MTK02] Only 50Mbps over-provisioning on half of the paths [ASS03]
12	Congestion – Recap Scalability Multiple plants eventually congest bandwidth Aggressive behaviors TCP expands to fill available bandwidth Aggressive flows Over-provisioning Does not deal with above
13	Quality-of-Service Problem Best-effort networks Congestion Metrics Implementation Application-middleware IP Data link
14	QoS Assurances So, what do we want from QoS? Worst-case Reasoning Reason about system timing Priorities make it easy Is it cross-traffic? Is it failures (e.g., network partition)? Failures as first class citizens Average-case Non-sense Exacerbated by TCP Tail E.g., 99-percentile Must be paired with middleware, fault-tolerance
15	Quality-of-Service Problem Best-effort networks Congestion Metrics Implementation Application-middleware IP Data link
16	Protocol Stack Dealing with complex systems Explicit structure allows identification, relationship of complex system’s pieces Layered reference model for discussion Modularization eases maintenance, updating of system Change of implementation of layer’s service transparent to rest of system E.g., change in gate procedure doesn’t affect rest of system
17	QoS
18	Middleware Between application and transport Libraries to provide advanced functionality Hide communication Applications Resource Discovery Remote Procedure Calls Security Interoperability (e.g., since Real-Time Corba) Scheduling, resource management, performance analysis Multicast Software development Simpler, faster State-of-the-art functionality Middleware over IP Wealth of libraries for IP Critical advantage of the Internet Protocol
19	Middleware and QoS End-point middleware Assume best-effort, unreliable network with no explicit QoS provision End-points compensate for vagaries Previous work (example) Voice-over-IP Replay buffer Address both congestion and transient failures E.g., routing transients
20	Sampling and Control [ACC03] Metrics Stability and tracking Delay characterization Heavy tail: Pr[delay>x]~x^-^a View delay spikes as transient failures Contingency control Used if no regular control received Recover from delay spikes
21	Evaluation Methodology Collect Round-Trip Times Case ↔ NASA (and other locations) Note: NASA is far from Case (in Internet distance) Simulate scalar plant with fast dynamics Use deadbeat control but compensate for nominal RTT Observations Base case diverges Contingency converges
22	QoS
23	QoS: from IP to data link QoS at data link E.g., priorities, token passing, VPN E.g., Ethernet QoS at IP Protocol IPv4: 4 basic Types-of-Service (ToS) IPv6: many flow ids, Classes-of-Service (CoS) QoS: from data link to IP Code-point mapping IP addresses
24	QoS in Multi-hop Networks Queuing Packets are forwarded from one network to another Packets are queued at routers (store-and-forward) Protect from cross-traffic cause congestion Per-Hop Behaviors (PHB) Behavior of individual router E.g., Forward before all other queued packets Maps into data link QoS Depends on Class-of-Service Class of service Assigned by source or shapers Imply PHB
25	QoS at IP Best effort IP IP is de facto a convergence layer Must deal with heterogeneous networks Architectural principle Difficult to deploy QoS at IP Hard to provide end-to-end QoS
26	QoS
27	QoS: LAN and WAN Data link Network layers Data link QoS independent of IP, transport, middleware Middleware and transport motivate IP Multiple possible solutions E.g., Ethernet switches with 802.1p, 802.1q Administrative Can do anything (e.g., RVSP) within any single administrative domain Need much cooperation with multiple administrative domains
28	Ethernet IEEE 802.1p Priority Three bits, recommended values 0-3=normal 4-7=high Multicast control Multicast packets forwarded after unicast packets with same priority IEEE 802.1q Objective: create Virtual LANs (VLAN) Cross-traffic, multicast control Locally administered
29	QoS: Summary Traffic isolation Cross-traffic can degrade performance Worst-case and probabilistic assurance Application and Middleware Adaptive Network Queueing, scheduling, forwarding Data link Priorities and beyond
30	Real-Time Protocol
31	Network Layer Convergence layer (IP) IP (Internet Protocol) IP supports multiple data links Multiple transports support IP Transport independent of data link Mix and match E.g., RTP/UDP over 802.11, Ethernet E.g., TCP over Infinet
32	Transport Layer Design choice TCP or RTP/UDP Three rules: Real-time data: RTP/UDP Sensor data Video, sound, I/O Control Bulk data: TCP Log download Software upgrades Coarse supervision Middleware rules E.g., RealAudio over UDP E.g., HTTP over TCP Usually, right choice made by middleware designer However, bad choices can propagate
33	Real-Time Transport RTP-RTCP RTP is a data transfer protocol RTCP is a companion control protocol Problems above solved Support for Feedback on connection quality Timing, synchronization End-point description, aggregation Frame boundaries Flexibility Application-level framing Custom-tailored reaction to congestion, packet losses, but No error detection, recovery Independent of underlying transport Can work over TCP Small header Multicast support
34	RTP Session Session identified by IP addresses Port numbers IP addresses Unicast or multicast RTP port number RTCP port number RTP port number + 1
35	RTP Packet Header Version (2 bits) Padding (1 bit) Extension (1 bit) CSRC count (4 bits) Marker (1 bit) Payload type (7 bits) Seqno (16 bits) Timestamp (32 bits) SSRC id (32 bits) CSRC ids (32 bits each) Payload Padding Padding Pad count (8 bits)
36	RTCP Control protocol Works in conjunction with RTP Functions QoS monitoring and congestion control Convey feedback on quality of data delivery and information of membership Modify transmission rates Diagnostic Source identification Session size estimation and scaling Session control E.g., user names to be displayed Reports Participants periodically send RTCP report packets Statistics, e.g., Number of packets sent Number of packets lost Inter-arrival jitter
37	Mixers Merge several media streams of the same types into one new stream Can reduce bandwidth consumption Mixer appears as new source Use new SSRC Put original SSRCs in CSRC list.
38	Translators Single media stream May convert encoding Protocol translation E.g., ATM↔IP Keep original SSRCs
39	Mixers and Translators
40	Implementation RTP Libraries Bell Laboratories (in C) http://www.bell-labs.com/topic/swdist/ JRTPLIB (in C++) http://lumumba.luc.ac.be/jori/jrtplib/jrtplib.html liveCaster and other tools http://live.sourceforge.net/ RTP Tools (rtpdump, rtpmon, rtpsend,…) http://www.cs.columbia.edu/IRT/software/rtptools/
41	RTPlib
42	RTPlib Get session information RTPMemberInfoGetStatus(); RTPMemberInfoGetRTPAddr(); RTPMemeberInfoGetPktCnt(); RTPMemberInfoGetUserInfo(); RTPSessionGetBandwidth(); RTPFindMember(); Advanced functions RTPSplitCompoundRTCP();
43	Conclusions Quality-of-Service Real-Time Protocols Web resources (next)
44	Web resources Network Control Systems At Case http://vincenzo.liberatore.org/NetBots/ Community-wide repository Under construction Network Research Lab http://vincenzo.liberatore.org/netlab/
45