DetNet Bounded Latency-04 drafu-fjnn-detnet-bounded-latency-04 Norman Finn, Jean-Yves Le Boudec, Ehsan Mohammadpour, Huawei EPFL EPFL Jiayi Zhang, János Farkas, Balázs Varga Huawei Ericsson Ericsson IETF 105 DetNet WG 7/16/2018 1 Montréal, 22 July, 2019
A reminder to new attendees … • DetNet is about an upper bound on end-to-end latency – not low average latency. • Bounded latency leads to the ability to compute exactly how many bufgers are required to achieve zero congestjon loss (and vice versa). • Feedback that slows down fmows to avoid congestjon is not an optjon for the applicatjon space of interest to DetNet. • Mathematjcally sound assurances can be given on latency and congestjon loss. 7/16/2018 2
Major changes from -03 to -04 • Sectjon 3 reorganized—the “reserve before use” paradigm applies to both the statjc and the dynamic latency computatjon problems. • All of the various supported queuing techniques have been made subsectjons of sectjon 6, “Queuing techniques”. • The difgerent queuing techniques have been given more equal atuentjon, some enhanced, some shortened. • Sectjon 8, “Parameters for the bounded latency model”, has been deleted. 7/16/2018 3
Clause 3 Flows are created by: 1. Confjgure the network. 2. Characterize the fmow. 3. Establish the path the fmow is to take. 4. Compute the ability of the network to handle the fmow and the suitability to the fmow’s requirements of the QoS ofgered, e.g. compute latency. • The Statjc latency computatjon: Recompute every fmow’s latency whenever any fmow is added or removed. • The Dynamic latency computatjon: Compute absolute worst-case latency once, when fmow is created. 5. If satjsfactory results, reserve the resources and give the sender permission to start. 7/16/2018 4
Clause 6: Queuing techniques 6.2 Preemptjon : The transmission of exactly one Ethernet frame can be suspended many tjmes, with critjcal frames transmitued in each gap. 6.3 Time-scheduled queuing : Each output queue is gated by a synchronized, rotatjng schedule set by management. 6.4 Asynchronous Traffjc Shaping : Hierarchical per-fmow and per-class shaping, with fewer than one queue per fmow. 6.5 IntServ : Hierarchical per-fmow and per-class shaping, without one queue per fmow. 6.6 Cyclic Queuing and Forwarding : Double- or triple- bufgering for each class on each port, with bufgers cycled in synchrony across network. 7/16/2018 5
6.6 Cyclic Queuing and Forwarding • Two-bufger version: Two bufgers per port. Input and output bufgers swap at the same moment, once every cycle, period T C . Small guard band to allow for transit and forwarding tjme. All relay nodes are synchronized and swap bufgers at the same moment. Cycle tjme T C > transit tjme + forwarding tjme + clock inaccuracy + max data transmit tjme. 7/16/2018 6
6.6 Cyclic Queuing and Forwarding • Two-bufger version: Two bufgers per port. Input and output bufgers swap at the same moment, once every cycle, period T C . Small guard band to allow for transit and forwarding tjme. All relay nodes are synchronized and swap bufgers at the same moment. Cycle tjme T C > transit tjme + forwarding tjme + clock inaccuracy + max data transmit tjme. TICK! 7/16/2018 7
6.6 Cyclic Queuing and Forwarding • Two-bufger version: Two bufgers per port. Input and output bufgers swap at the same moment, once every cycle, period T C . Small guard band to allow for transit and forwarding tjme. All relay nodes are synchronized and swap bufgers at the same moment. Cycle tjme T C > transit tjme + forwarding tjme + clock inaccuracy + max data transmit tjme. TICK! 7/16/2018 8
6.6 Cyclic Queuing and Forwarding • Three-bufger version: Three bufgers per port. Same as two- bufger version, but input bufger swap is out-of-phase with output bufger swap to allow for arbitrary link delay. 7/16/2018 7/16/2018 9 9
6.6 Cyclic Queuing and Forwarding • Three-bufger version: Three bufgers per port. Same as two- bufger version, but input bufger swap is out-of-phase with output bufger swap to allow for arbitrary link delay. TICK! 7/16/2018 7/16/2018 10 10
6.6 Cyclic Queuing and Forwarding • Three-bufger version: Three bufgers per port. Same as two- bufger version, but input bufger swap is out-of-phase with output bufger swap to allow for arbitrary link delay. TICK! 7/16/2018 7/16/2018 11 11
6.6 Cyclic Queuing and Forwarding • Three-bufger version: Three bufgers per port. Same as two- bufger version, but input bufger swap is out-of-phase with output bufger swap to allow for arbitrary link delay. TICK! 7/16/2018 7/16/2018 12 12
6.6 Cyclic Queuing and Forwarding • Three-bufger version: Three bufgers per port. Same as two- bufger version, but input bufger swap is out-of-phase with output bufger swap to allow for arbitrary link delay. TICK! 7/16/2018 7/16/2018 13 13
6.6 Cyclic Queuing and Forwarding • Three-bufger version: Three bufgers per port. Same as two- bufger version, but input bufger swap is out-of-phase with output bufger swap to allow for arbitrary link delay. TICK! 7/16/2018 7/16/2018 14 14
6.6 Cyclic Queuing and Forwarding • Three-bufger version: Three bufgers per port. Same as two- bufger version, but input bufger swap is out-of-phase with output bufger swap to allow for arbitrary link delay. TICK! 7/16/2018 7/16/2018 15 15
6.6 Cyclic Queuing and Forwarding • Time-based CQF is defjned in IEEE 802.1 standards. • Packet-marker based CQF is suggested in private DetNet drafus. • CQF can be operated at multjple frequencies on one port to serve more than one Class of Service (bandwidth/latency range): Slow Medium Fast 7/16/2018 16
Summary* Technique Latency Overprovisioning Handles State required Time sync computatjon necessary predictably per-hop required bursty fmows 6.3 Time- Statjc Small Yes Per class schedule Yes scheduled NP hard 6.4 IntServ Statjc (recompute Small No Per-fmow state, No all fmows on any per-fmow queue change) 6.5 Time-Aware Statjc (recompute Small No Per-fmow state, No Shaping all fmows on any per-port-pair queue change) 6.6 Cyclic Queuing Dynamic (trivial More No None Yes & Forwarding additjon) * This table is a generalizatjon. There are many factors that can mitjgate the difgerences. Other queuing schemes have been proposed that make other trade-ofgs.
Final steps… • Refjning the terminology to conform DetNet. • Using DetNet terminology and terms. • Formal delay analysis of CQF. • Per-node bufger size calculatjon. • Consistency check with the other WG drafus. 24/07/2019 18
QUESTION • Are we ready for adoptjon? 7/16/2018 19
Thank you 7/16/2018 20
Recommend
More recommend
