04:10 <kc8apf> adamgreig: 1588 over commodity switches is fine until the switch latency variability becomes significant.

08:02 <adamgreig> kc8apf: makes sense i guess, any idea roughly when that happens or how bad it can be?

08:03 <adamgreig> it has occurred to me that the current "client sends delay_req immediately upon seeing follow_up" has obvious problems when there's many clients, but should be easy enough to spread them over the sync period (in my case)

12:02 <cpresser> Win 17

15:23 <kc8apf> adamgreig: NTP can easily get you down to 10s of ms. On an ARM Linux system, a local GPS receiver with 1PPS got me down to 10s of usec.

15:24 <kc8apf> 1588 support in switches is mostly about a deterministic fast path. Without that, you get variability based on queuing.

15:24 <adamgreig> more so than the switch itself having a boundary/transparent clock?

15:24 <adamgreig> i'm looking to synchronise clocks on 10s-100s of devices on a single ethernet segment to better than 1us

15:25 <adamgreig> i managed to bodge together an stm32-based test last night which is getting <100ns without heavy loading on the switch, just with a cheap netgear desktop thing

15:25 <adamgreig> but not sure how well it will scale with switch load or more devices etc

15:26 <adamgreig> i had assumed ptp switches would also correct the ptp packets for time in transit through the switch

15:26 <kc8apf> As port count grows, buffer size will also grow. If 1588 is the only traffic, it might work. Throw in anything else and you'll get loss of variability.

15:27 <adamgreig> most of the nodes are likely to be sending a lot of traffic but not receiving much except 1588

15:28 <adamgreig> not sure what a switch buffer architecture looks like though

15:28 <kc8apf> That will still cause queuing

15:28 <adamgreig> will heavy outbound traffic affect inbound traffic, or is it just a question of total switch traffic?

15:28 <kc8apf> At 100s of nodes, you are looking at multiple 48 port switches anyway

15:29 <adamgreig> mostly 24port switches for PoE power budget, but probably should say 10s-100 rather than 100s

15:30 <kc8apf> Depends on the architecture. Anything based on broadcom's big chips use a shared memory pool that each port pulls from.

15:30 <kc8apf> If you're looking at that big of a switch, just buy one with 1588. It's common enough

15:30 <adamgreig> right, so lots of unrelated traffic on other ports would still cause serious queueing for ptp? that's a shame

15:31 <adamgreig> I guess I was looking for switches that were 1588 clocks rather than just having a 1588 fast path?

15:32 <kc8apf> If you build your 1588 client well, it should eliminate any deterministic delay

15:33 <kc8apf> I highly suggest looking at the chrony source for inspiration. It's well written and fairly easy to understand.

15:34 <adamgreig> this is going to be gateware in the end, but i'll have a look

15:35 <adamgreig> chrony is just ntp right?

15:35 <kc8apf> Nope. It does NTP but also 1pps, ptp, etc.

15:35 <kc8apf> Technically ntpd can do those as well but it's much hackier.

15:37 <adamgreig> oh cool, didn't know chrony did ptp

15:38 <kc8apf> Keep in mind that you can combine multiple sources as well. That's how 1pps works. The actual 1pps edge only tells you when the second change occurred with only provides relative time. You combine it with another source to get absolute time.

19:30 <mossmann> adamgreig: You might be interested in White Rabbit: https://white-rabbit.web.cern.ch/

21:25 <mithro> How would people draw a flip flop with asynchronous reset verses a flip flop with an asynchronous reset?

21:33 <tnt> mithro: I'm not aware of a standard way to represent that

21:34 <mithro> Seems it is pretty common to use PRE / CLR verses SET / RST