SLIDE 1 Multiple Cyclic Queuing and Forwarding
(slides to accompany df-finn-multiple-CQF-0919-v01)
Norman Finn Huawei Technologies Co. Ltd nfinn@nfinnconsulting.com df-finn-multiple-CQF-slides-0919-v01
SLIDE 2 Why this paper?
Cyclic Queuing and Forwarding is perceived as having major flaws, which prevent its adoption:
- It is difficult to pick a cycle time:
- A long cycle time produces long end-to-end latency.
- A short cycle time severely restricts the number of flows handled.
- No one cycle time serves more than a narrow range of requirements.
- Having 2-buffers per hop precludes long links.
With a little imagination, these issues can be solved without adding anything to IEEE 802.1Q, so the solutions are suitable for P802.1DF.
SLIDE 3
Step 1: You can run with 3 buffers, not 2
! With 2 buffers, you change input buffers at the same moment that you change output buffers.
! Link delay and forwarding delay have to be subtracted from the useful bandwidth, to prevent filling and emptying a buffer at the same time.
! If you have 3 buffers, then you can change to the next input buffer at a different phase of the cycle from when you change to the next output buffer.
! Each buffer spends part of a cycle idle; one cycle per hop is added to the latency. ! But, the link delay and forwarding delay are no subtracted from the bandwidth..
SLIDE 4 Two buffers
Two buffers per port. Input and output buffers swap at the same moment, once every cycle, period TC. Small guard band to allow for link delay and forwarding delayu. All bridges are synchronized and swap buffers at the same moment. Cycle time TC > transit time + forwarding time + clock inaccuracy + max data transmit time.
7/16/2018
4
SLIDE 5 Two buffers
7/16/2018
5
TICK!
SLIDE 6 Two buffers
7/16/2018
6
TICK!
SLIDE 7 Three buffers
Input buffer swap is out-of-phase with output buffer swap to allow for arbitrary link delay. No guard band needed for link/forwarding delay. Each buffer cycles through four states: filling, full, draining, empty
7/16/2018
7
7/16/2018 7
SLIDE 8 Three buffers
7/16/2018
8
7/16/2018 8
TICK!
SLIDE 9 Three buffers
7/16/2018
9
7/16/2018 9
TICK!
SLIDE 10 Three buffers
7/16/2018
10
7/16/2018 10
TICK!
SLIDE 11 Three buffers
7/16/2018
11
7/16/2018 11
TICK!
SLIDE 12 Three buffers
7/16/2018
12
7/16/2018 12
TICK!
SLIDE 13 Three buffers
7/16/2018
13
7/16/2018 13
TICK!
SLIDE 14 N buffers
Any number of buffers per port. Useful for delay matching on different paths.
7/16/2018
14
7/16/2018 14
SLIDE 15 The number of buffers used by a stream
! Different output ports on one system can have different numbers
! Different input ports feeding the same output port can each use a different number of the buffers on that output port. ! One Stream can use a different number of buffers at each output port along the path.
SLIDE 16 Timeline 2: Assigning frames to output buffer in Node B 2-buffer: discard 3-buffer: send to buffer a T=0.3 T=1.3 Timeline 3: Storing frames in output buffers in Node B T=0.5 T=1.5 T=3.5 T=0 T=1 T=2 T=3 2-buffer: send to buffer c 3-buffer: send to buffer a
NODE B
T=2.5
Minimum forwarding delay Maximum forwarding delay Forwarding delay variation
Timeline 4: Transmitting frames from Node B
Slope of maximum forwarding delay TW wait time
T=0 T=1 T=2 T=3 T=2.3 T=3.3 Timeline 1: Transmitting frames from Node A minimum link delay = 1.3TC
max interference time TI variation time TV
- utput continues (3-buffer)
NODE A NODE B NODE B
dead buffer time TB
- utput continues (2-buffer)
allocable to Streams TA+TP
df-finn-multiple-CQF-0919-v01 Figure 1 Reference timelines for time-based CQF
SLIDE 17
Example of Multiple cycle times on one output port
! Four cycle times at four priorities 3-6. ! 1 priority 3 cycle, buffers h, i (not shown) ! 3 priority 4 cycles, buffers f, g ! 6 priority 5 cycles, buffers d, e ! 24 priority 6 cycles, buffers a, b, and c.
SLIDE 18
Best effort 4 5 6 Time -->
Transmission order
! The total bandwidth allocated over all CQF priorities cannot exceed the link bandwidth. ! Draining of CQF buffers are by strict priority, fastest cycles first ! As long as every cycle is an integer multiple of the next-faster cycle, every buffer will empty before the end of its cycle.
SLIDE 19
More ways to improve the options available to CQF.
! Preemption:
!Lowers the interference from lower-priority queues, especially best-effort !CQF queues can be preemptable, because the preemption penalty is bounded.
! Packing varied frame sizes into cycles can require overprovision:
!“Sausage making,” packing the customers’ packets into fixed-sized encapsulations, requires much less overprovision, and can be done at the edges of a Service Provider network.
! In multiple CQF, deliberate overprovision of a Stream (assigning it to a too-fast cycle) reduces its latency.
SLIDE 20
The multiple-CQF value proposition
! CQF, as described here, requires network time synchronization. ! CQF requires time-aware state machines on input/output ports. but, ! Except, perhaps, for the ingress bridge, CQF requires no per- Stream configuration, no per-Stream state machines, no per- Stream knowledge at any node in the network. ! CQF allows a network controller to perform admission control with no per-flow conversations with any (except perhaps the ingress) nodes in the network.
SLIDE 21 To come …
There are many more improvements that can be made to TSN, many
- f which require no new standards. I hope to supply more papers
and presentations in this vein.
SLIDE 22
Thank you
