Let’s dive a little deeper into Time-Aware Shapers, specified in the IEEE 802.1Qbv standard. For a light introduction go read my intro to TSN.
Gates
The shapers are applied on egress and each shaper has a list of time slots,
where each time slot specifies a duration, operation, and gate_mask. The
duration represents how long the time slot is. The gate_mask specifies
which egress queues are open during that time. Multiple queues can be
open at the same time, at which point it treats it as regular Strict Priority
between those queues, i.e., higher priority goes first. operation will be
covered later. Time-Aware Shapers can have a theoretically infinite number of
these time slots, also known as gates, but in practice there will of course
be a limit to what both the software and hardware can support. The sum of the
gates’ durations defines the cycle time. When it has iterated through all the
gates it restarts with the first gate again. Below is an example of a schedule
with two gates, one which opens priorities 0-3 and one which opens priorities
4-7.
|---------|---------|---------|---------|
| 0,1,2,3 | 4,5,6,7 | 0,1,2,3 | 4,5,6,7 | ...
|---------|---------|---------|---------|
You should always make sure that all possible priorities in your system should be able to send at least during one gate. Otherwise it will take up space in the queue and never be sent. You can of course use only priorities 0-2 if you wish, but then you must make sure that no priority is ever mapped to the rest. The priority is usually the PCP or DSCP value, but how you map that to the internal priority of the switch is up to you. Since DSCP can have 64 different values you already need some sort of reduced mapping to the 8 queues used by the shaper.
The lower limit of what defines reasonable durations is that frames need to still be able to pass through with little interruption. Too short and it will start causing more problems that it solves. With too short gates it could result in frames either not fitting in the gate and start encroaching on the next time slot, or being preempted a lot if used with Frame Preemption.
The maximum latency for your requirements defines the upper limit of a single gate’s duration A low-priority gate cannot have a duration longer than the maximum latency, otherwise the higher priorities has to wait too long for their turn. The total cycle time has no theoretical upper limit, but at some point it will be more worth to have it cycle back to the start instead of defining more gates.
All gates have a guard band at the end of their duration. During the guard band no new frames can be sent, it only allows already started frames to finish. This is to avoid frames encroaching on the next time slot. To completely avoid it you can set the guard band to the maximum size of a frame, 1518 bytes. This means that if the frame started sending it will be guaranteed to finish before the gate closes. However, this could end up wasting a lot of time when a frame finishes very early into the guard band and it has to idle until the next gate.
|------------------|-------------|
| Gate 1 | GB | Gate 2 | GB |
|------------------|-------------|
Frame Preemption
Frame Preemption can be used in conjunction with Time-Aware Shapers to make them more efficient. Frame preemption was covered briefly in the TSN introduction post. A quick recap. It requires frames to be classified as express or non-express frames. It allows non-express frames to be interrupted by express frames to allow the express frame to send immediately.
When used with Time-Aware Shaper you apply the different Frame Preemption
options SetGateStates (S), Set-And-Hold-MAC (H), and Set-And-Release-MAC
(R) to the different gates. S does nothing with respect to Frame Preemption,
but is an indicator that the guard band needs to be longer before it starts. H
makes sure any frames from previous gate are preempted when the gate starts. R
allows non-express frames to send again. H allows the guard bands to be
reduced drastically because now they can preempt a frame when reached. It only
needs to add the extra CRC checksum and then it can move on to the next gate.
Shorter guard bands enables better bandwidth usage.
If you have multiple express gates (H) in a row then it won’t be able to
preempt in preparation for the next gate, but then you also have bad design :).