IEEE 802.1Qbv: Time Aware Shaping for Lowest Latency in Ethernet Networks

1. Joint IEEE-SA and ITU Workshop on Ethernet Overview of IEEE 802.1QbvTime Aware Shaping Don Pannell, Principal Systems Architect Marvell Semiconductor

2. Need/Desire/Goal of Qbv • Get The Lowest Latency Possible – Any Way Possible • Want Many Long, ~32 hop, Daisy Chains • Small Bursts of frames at known regular intervals (e.g., a 40 uSec long burst of data every 125 uSec) • Willing to Engineer Network Segments to meet this goal • Non-Engineered (i.e., Consumer) Networks will not be able to depend on this very low latency as it can’t be guaranteed in their Networks • The Network Structure and Usage will have to be Engineered, Managed and Controlled

3. How Fast Can A Bridge Go? • Total time for the Critical frame is: • Internal delay of the Bridge • Plus the time to transmit the max size interfering frame • Plus the time to get the bits down a 100 meter cable • Tx of max frame overlaps the Rx of the Critical frame • Detailed analysis is shown in http://www.ieee802.org/1/files/public/docs2011/new-pannell-latency-options-0311-v1.pdf • The max size interfering frame is the determining factor • How does this improve if the interfering frame wasn’t there?

4. How Fast If No Interference? • Totaltime for the Critical frame is: • Internal delay of the Bridge • Plus the time to transmit the largest frame size of the Critical stream • Plus the time to get the bits down a 100 meter cable • Now the max size of the Critical frame is the determining factor • This is due to the Store & Forward nature of most Bridges • Since most Critical streams use small frames this is a great improvement • And the Critical stream frame size can be part of the ‘Engineering of the Network’

5. Interfering Frames are the Problem • Some GE speed examples with numbers: • The Bridge Latency with Interfering Frame is: • Equal to the Size of Interfering frame + Bridge Delay + Cable Delay (max size interfering frame = 1522 bytes) • With Max Size interfering frame this is 13.898 uSec • The Bridge Latency without an Interfering Frame is: • Equal to the Size of AVB/TSN frame + Bridge Delay + Cable Delay • With a 300 byte AVB/TSN frame this is 4.122 uSec • Can we get rid of the Interfering Frames to get the better latency?

6. Qbv – Time Aware Shaper • Qbv takes advantage of the target low latency data pattern • i.e., That they are typically Small Bursts of frames at known regular intervals (for example: a 40 uSec long burst of data every 125 uSec) • Then use this information to delay the start of non-Critical frames just before the start of the Burst Window • This insures the egress port is idle so the Critical burst is not interfered – All interference is removed! • Smart designs can allow non-Critical frames that fit to use the available bandwidth

7. Inside an Qbv Bridge (example) • Qbv Time Progression – Fig 1 • At Bridge t0-16.000 uSec before the start of the Burst Window the Green AVB Class B frames are being Shaped (gated) by Qav and can’t Transmit • So the Red Max size non-AVB High Priority frame ‘n’ can start

8. Inside an Qbv Bridge (example) • Qbv Time Progression – Fig 2 • At Bridge t0-3.664 uSec before the start of the Burst Window the interfering Red Non-AVB frame is done • Now the Green AVB Class B frames are available for transmit with enough credits to burst two frames 16.000-12.336= 300+20 bytes = 2.560

9. Inside an Qbv Bridge (example) • Qbv Time Progression – Fig 3 • At Bridge t0-1.104 uSec before the start of the Burst Window the 1st Green AVB Class B frame is done • Now the next Green AVB Class B frame has credit to go, but it can’t because there is not enough time before t0 - the start of the Burst Window nor can the Red ‘m’ frame • But the 64 byte low priority Yellow non-AVB frame can go and does 3.664-2.560= 64+20 bytes = 672

10. Inside an Qbv Bridge (example) • Qbv Time Progression – Fig 4 • At Bridge t0-0.432 uSec before the start of the Burst Window the 64 byte Yellow frame is done • The next Green AVB Class B frame has credit to go, but it still can’t because there is not enough time before t0 (its credits are actually increasing) nor can the Red ‘m’ frame • The next low priority 64 byte frame can’t go either – not enough time 1.104-0.672= 64+20 bytes = 672

11. Inside an Qbv Bridge (example) • Qbv Time Progression – Fig 5 • At Bridge t0 - the start of the Burst Windowthe port is idle so the newly arrived Blue Critical frames are allowed to egress without any interference! • The burst of Blue frames will continue as long as Qbv leaves it queue open for transmission as they are the highest priority queue

12. Qbv – With Cut Through • Cut-through bridges generally don’t help normal network performance due to the low percentage of improved latency and that this improvement cannot be guaranteed • With Qbv the improved latency can be guaranteed • Cut-through only works when the target ports are idle and Qbv does exactly that – thus the guarantee • With Cut-through the Bridge latency is:

13. Latency By The Numbers • Non-Qbv Bridge Latency is 13.898 uSec • Due to the Max Size interfering frame • Store & Forward Qbv Bridge Latency is 4.122 uSec • Assuming a 300 byte maximum size AVB/TSN frame • This number will go up with increasing Critical frame sizes • Cut Through Qbv Bridge Latency is 2.074 uSec! • Assuming a 64 byte Cut Through Point • This number does NOT change due to frame size! • And this performance can be Guaranteed!

14. Latency By The Numbers • Qbv supports the lowest latency possible • But only if the network is ‘Engineered’ • It won’t work for all topologies • And proper configuration of the timing ‘gates’ is not easy • The Qbv Standard will support Control of the timing ‘gates’ – but will not figure the timing out • 3rd Party tools will be required to compute the ‘gate’ timing • Multiple timing ‘gate’ windows will be supported per port with each window having nSec configuration resolution • TSN Bridges already have network timing information by their support of IEEE 802.1AS

15. Questions?