Quality of Service (QoS) One definition: Network capability to provide a non-default service to a subset of the aggregate traffic (could be better than default, could be worse than default) Dominant service class in use today: “best effort”

Use of QoS to deal with network congestion issues Can’t you just increase the bandwidth and eliminate need for QoS? Adding bandwidth doesn’t resolve jitter or provide traffic prioritization

Major paper in 1992 SIGCOMM Proceedings: first serious description of service classes – describes two mechanisms: Token bucket filter: enforcing treatment for a particular flow Weighted fair queuing: devices bandwidth across queues of traffic based on weights, ensuring that each traffic flow gets a fair share

Layer 2 Class of Service (CoS) IEEE 802.1p standard for prioritization (“Traffic Class Expediting”) Use of 3-bit “user priority field within the 802.1Q tag 802.1p capable NICs, switches or routers can tag packets; transit switches or routers implement the priority through prioritized queuing

802.1Q tag frame format TPID = 8100 When a frame has EtherType code of 8100, we know there’s an 802.1Q tag Priority: 3 bits for priority as per 802.1p CFI: Ethernet or Token Ring (0 for Ethernet) VID: VLAN ID – 12 bits: 4094 VLANs (0 and 4095 reserved)

802.1p Service Classes The 8 classes/levels of priority available from the 3 bits are assigned as follows: 000 (0) - Routine 001 (1) - Priority 010 (2) - Immediate 011 (3) - Flash 100 (4) - Flash Override 101 (5) - Critical 110 (6) - Internetwork Control 111 (7) - Network Control Some switches have four queues per port; some have eight Weighted Round Robin usually used to service queues (alternative: strict queuing)

Layer 3 CoS/ToS: IP Precedence Type of Service field in IP header When used, IP Precedence has traditionally used first three bits of TOS field for 8 possible value Guess how the 8 possible precedence values have been defined?

Limitations of IP Precedence IP-Precedence scheme allows only specification of relative priority of a packet; no provisions to specify different drop precedence for packets of a certain priority. Example: you may want to specify both HTTP and Telnet as high-priority. However, if congestion occurs, you want Telnet packets to be dropped, before HTTP (or vice versa). 3 bits restrict the number of possible priority classes to eight. The Network Control and Internetwork Control classes are usually reserved for router-generated packets, necessary for the health of the network. This cuts down usable classes for production traffic to 6. IP-Precedence is not always implemented consistently by network vendors today.

Internet/IETF QoS Efforts 1994 – IETF began work on an Integrated Services (IntServ) architecture: Flow = stream of packets with common Source Address, Destination Address and port number Requires router to maintain state information on each flow; router determines what flows get what resources based on available capacity

IntServ components Traffic classes Traffic control best effort controlled load (‘best-effort like’ w/o congestion) guaranteed service (real-time with delay bounds) Traffic control admission control packet classifier packet scheduler

IntServ components (cont.) Setup protocol: RSVP “Path” msg from source to destination collects information along the path; the destination gauges what the network can support, then generates a “Resv” msg If routers along the path have sufficient capacity, then resources back to the receiver are reserved for that flow; otherwise, RSVP error messages are generated and returned to the receiver Reservation state is maintained until the RSVP “Path” and “Resv” messages stop coming

IntServ/RSVP problems Scalability (processing of every individual flow on core Internet routers) Lack of policy control mechanisms

What is the service? There are two camps: “Better best effort” — ISPs want finer control of relative bandwidth allocation, particularly under heavy load (Implementation in terms of drop- preference or weighted- round- robin). “Virtual leased line” — Users want an end- to- end absolute bandwidth allocation, independent of other traffic (Implementation in terms of priority queuing and strict policing). The IETF is more vocal on the former but there is demand for both.

What are the target applications? Bad question. In 1978, the answer was remote job entry. In 1988, email/ ftp. In 1999, web. This too will change. IP/ TCP/ UDP/ IGMP/ OSPF/ BGP work for any application. Differentiated services must too. What are the requirements of today’s applications?

Scaling A differentiated services mechanism must work at the scale of the Internet (e. g., millions of networks) and at the full range of speeds of the Internet (e. g., Gb/ s links). The Internet is growing at 20% per month. To get that kind of scaling the design must: push all the state to the edges, and force all per- conversation work (e. g., shaping, policing) to the edges.

Scaling (cont.) Edge- only state suggests that special/ normal service indication must be carried in the packet. Administrative diversity and high speed forwarding both argue for very simple semantics on that indication. E. g., a few bits of special/ normal. No state in center means it sees only aggregates (potential fairness problems).

Differentiated Services (DiffServ) See http://www.ietf.org/html.charters/diffserv-charter.html Instead of maintaining individual flows on all routers, flows are aggregated into an aggregate flow that receives “treatment” (per class or per service state) Service classes are identified, packet is marked as belonging to a particular service, sent on its way; routers in path examine header to determine treatment

Necessary DiffServ functions Admission control: ability of network to refuse customers when demand exceeds capacity Packet scheduling: method for treating different customers’ data differently as needed Traffic classification: ability to sort streams into “substreams” that receive different treatments Policies and rules for allocating the network’s resources

DiffServ from the router’s perspective define packet treatment classes allocate an adquate amount of resources for each class sort all incoming packets into their corresponding classes

DiffServ components DS-field: uses 6 bits in the TOS field in IPv4/traffic class field in IPv6; denotes service the packet should receive (6 bits referred to as the DSCP: DiffServ Code Point) PHB (per-hop behavior): defines the service at each hop; may be relative (compared to other PHBs) or absolute (in bandwidth or delay terms) BA (behavior aggregate): group of packets with the same DSCP (PHB is applied to each BA)

Note error in above: 101-Immediate should be 010-Immediate

Proposed DiffServ services Expedited Forwarding - “virtual leased line”; low delay, loss and jitter; provides a peak rate/ceiling (DSCP: 101110) Assured Forwarding - emulates a lightly loaded network (“drop me last”); provides a “floor” rate Default - usual “best effort” service (DSCP: 000000)

DiffServ AF Codepoint Table Drop Probability Class #1 Class #2 Class #3 Class #4 Low Drop Probability (AF11) 001010 (AF21) 010010 (AF31) 011010 (AF41) 100010 Med Drop Probability (AF12) 001100 (AF22) 010100 (AF32) 011100 (AF42) 100100 High Drop Probability (AF13) 001110 (AF23) 010110 (AF33) 011110 (AF43) 100110

Possible DiffServ Class Provisioning Traffic Class DSCP IP Routing CS6 Interactive Voice EF Interactive Video AF41 Streaming Video CS4 Telephony Signaling CS3 Transactional/Interactive AF21 Network Management CS2 Bulk Data AF11 Scavenger CS1 Best Effort

Fun Issues and Questions How will 802.1p, DiffServ and RSVP interoperate? Not enough to just identify “audio/video” packets (relative to others) and let them thru first. May have more “audio/video” packets than bandwidth available (LOTS of people running audio/video apps). Who decides WHOSE audio/video packets? How do we manage user expectations? How do users view QoS? How do they invoke it? How do they know that they have it?

Fun Issues and Questions (cont.) What to do with incoming traffic? What do we do with flows currently in progress when a higher priority flow requests access? Do we allow preemption? What about the problem of multiple sources sending data to a single destination, which then ends up oversubscribed?

Fun Issues and Questions (cont.) How do we prevent starvation of “best-effort” traffic? How do you manage QoS policies? How do we decide who gets what? Are DiffServ PHBs like a hall pass in school to go to the bathroom? If so, who decides who has to go the worst? Does it come back to pricing?

Qbone Testbed work: Two models: http://qbone.internet2.edu Premium: the goal of offering as close to a virtual leased line service model as is possible over IP Scavenger: allows for marking traffic for potentially degraded treatment at congested downstream interfaces Note: “Due to a medium-large set of intractable deployment problems, the Internet2 QBone Premium Service initiative has been suspended indefinitely. Internet2 QoS design and deployment efforts are now focused on ‘non-elevated’ forms of QoS like QBone Scavenger Service that require no policing, no reservations, and no admission control.”