Integrated Service Mappings on IEEE 802 Networks
Extracted elements (23)
A delay bound has no scope outside a single L2 domain. The end-to-end delay for a flow must be computed by summing per-domain values using standard RSVP path mechanisms; the L2 domain exports its contribution as C and D terms to L3 devices.
A universal, statically defined mapping of delay classes was considered but rejected as too rigid to map the wide range of user requirements onto the limited number of 802.1D classes. Instead, clients query the network for the appropriate user_priority for each flow, allowing the network to factor in current topology, load, and admitted flows.
Class 1 ('less than best effort') is designated for penalty-tagging flows that exceed their Tspec allocation, providing isolation from in-contract flows and enabling preferential discard by overloaded downstream devices. Tagging only individual out-of-Tspec packets is discouraged because it causes packet reordering.
Classes 4, 5, and 6 are recommended for delay-sensitive, admission-controlled flows because on 802.1D switches with only two-level strict priority queuing (default split 0–3 / 4–7), all three classes receive 'high priority' treatment.
Network Control traffic is placed at user_priority 7 to protect routing updates and network management. However, because this competes with QoS application traffic, SBMs performing admission control must include an estimate of Network Control bandwidth in their calculations.
Per-flow filtering at the switch level becomes prohibitively expensive with increasing switch speed, making devices with such capability nearly as complex as IP routers. The document adopts an 'aggregated flow' approach using IEEE 802.1D user_priority traffic classes instead.
Quantified delay bounds can only be realized in combination with admission control; priority scheduling alone provides only relative ordering and cannot enforce absolute delay or loss bounds for individual flows. The admission control algorithm determines which Int-Serv service level (Controlled Load vs. Guaranteed) is actually offered.
The user_priority mapping table is held by switches (and optionally routers) rather than end-station hosts to simplify upgrades: when new services or queuing algorithms are deployed, only network elements need updated mappings, not all end stations.
Controlled Load service has less stringent path requirements than Guaranteed Service, but switches must still provide differential queuing and buffering for CL flows over best effort. Where non-QoS-capable devices exist, policing must be approximated at network edges.
IEEE 802.1D-1998 switches using only default user_priority behavior cannot police flows internally; policing for IEEE 802 networks is therefore implemented at layer-3 edge devices, making it inherently approximate.
Providing Guaranteed Service requires that *all* switching elements along the path participate in special treatment for the flow's packets. Even a single switch transmitting onto a shared or half-duplex LAN segment breaks Guaranteed Service guarantees entirely.
The default mapping table in Table 1 is described as useful but preliminary; it may be refined in future editions. Network managers may override defaults via management control, but all switches within a domain must agree on mappings to some extent.
A merged RSVP reservation MUST receive at least as good a service as the best of its component reservations. Merging reservations that would demote or promote a flow into a different traffic class is declared illegal; an RSVP/SBM ResvErr is returned instead.
All data packets of a flow MUST be assigned to the same traffic class for the duration of the flow. This ensures packet ordering between adjacent Layer 3 routers and avoids reordering problems.
For Guaranteed Service, the C and D error terms SHOULD overestimate delay rather than underestimate it, so that users receive better performance than requested when approximations are used.
MTU and minimum packet size information MUST be reported accurately. 'Break bits' MUST be set correctly, both the overall QoS-support bit and per-service-type bits.
A Best Effort-only network element sets the 'break bit' defined in RFC 2210 (RSVP with Integrated Services) to indicate it cannot provide a specific QoS service. No explicit characterization parameters need to be computed.
A network element implementing Controlled Load must determine receive and transmit resource requirements (e.g., buffering) for each candidate flow to support admission control. Controlled Load exports no service-specific characterization parameters; internal estimates must account for statistical aggregation of CL flows into 802 traffic classes.
A network element supporting Guaranteed Service must compute and export: (1) constant delay bound 'C', (2) rate-proportional delay bound 'D', and (3) receive/transmit resource requirements (buffering, bandwidth) used by admission control.
General characterization parameters required of all service types include: ingress link, egress link MTUs/framing overhead/minimum packet sizes, available path bandwidth (updated hop-by-hop), and minimum latency. MTU and minimum packet size must be reported accurately; bandwidth may be loosely estimated.
The IEEE 802.1Q user_priority field carries a traffic class indicator within data frames. It takes values 0–7, with Table 1 providing the default mapping: 0 = Best Effort, 1 = less than Best Effort, 2–3 = reserved, 4 = Delay Sensitive (no bound), 5 = Delay Sensitive (100 ms), 6 = Delay Sensitive (10 ms), 7 = Network Control.
The SBM TCLASS object is inserted into RESV messages to carry the chosen user_priority value. When a second RESV for the same flow reaches a switch with an existing reservation, the TCLASS values are compared; if they differ, a ResvErr is returned to the originating L3 device.
QoS mechanisms open the possibility for denial-of-service and theft-of-service attacks. Any use of these service mappings assumes that all requests for service are authenticated appropriately. No new security issues beyond those in companion documents RFC 2816 and RFC 2814 are introduced.