Actually, we don't discuss metrics at all (yet). I think we should, and AFAIR this overlaps with the draft pushed by Jan and the other Cisco folks.
If we are aiming for "competition benchmark," another important point is that we must define what the expected behavior of the device is. To give an (absurd) example, I can trivially minimize my energy usage by dropping all packets, or not activating the ports, or not even turning on. Less trivially, if we want to foster the exploration of latency VS power trade-off, where does one put the upper-bar for latency? Or packet drops?
Debatable. What makes a characterization "correct"? The max power is correct and it is important for dimensioning supply circuits. But it is not sufficient for power optimization.
Maybe a good way to get a realistic scenario would be to use a snapshot of an actual router deployed somewhere right now. We could get traffic and RIB/FIB state (maybe transceiver models, too) from a couple of devices and use that. Plus a couple of synthetic extreme cases, a snake-test, etc. That could be a good start.
Actually, pretty much everything matters. The real challenge is to have a clean way of specifying and replaying the traffic trace you want to use. PCAPs don't scale well to 32*100G interfaces :-/
Sure, everything must be reported. To me though, the important thing I'd hope to get from the document is not so much what I must report on, but how this reporting should look like.
I don't understand the "should be proportional to its traffic load" part. We know this is not the case, and will never be. Or do you mean just the dynamic power?
I'm not sure how to best comment on parts of the draft... here is a try.
https://github.com/cpignata/powerbench/blob/ada3322da647a2b83ab4f3a97c6b1d12ada550f7/draft-cprjgf-bmwg-powerbench.xml#L85
Actually, we don't discuss metrics at all (yet). I think we should, and AFAIR this overlaps with the draft pushed by Jan and the other Cisco folks.
If we are aiming for "competition benchmark," another important point is that we must define what the expected behavior of the device is. To give an (absurd) example, I can trivially minimize my energy usage by dropping all packets, or not activating the ports, or not even turning on. Less trivially, if we want to foster the exploration of latency VS power trade-off, where does one put the upper-bar for latency? Or packet drops?
https://github.com/cpignata/powerbench/blob/ada3322da647a2b83ab4f3a97c6b1d12ada550f7/draft-cprjgf-bmwg-powerbench.xml#L92-L94
Debatable. What makes a characterization "correct"? The max power is correct and it is important for dimensioning supply circuits. But it is not sufficient for power optimization.
https://github.com/cpignata/powerbench/blob/ada3322da647a2b83ab4f3a97c6b1d12ada550f7/draft-cprjgf-bmwg-powerbench.xml#L136
Maybe a good way to get a realistic scenario would be to use a snapshot of an actual router deployed somewhere right now. We could get traffic and RIB/FIB state (maybe transceiver models, too) from a couple of devices and use that. Plus a couple of synthetic extreme cases, a snake-test, etc. That could be a good start.
https://github.com/cpignata/powerbench/blob/ada3322da647a2b83ab4f3a97c6b1d12ada550f7/draft-cprjgf-bmwg-powerbench.xml#L174-L176
I'm not sure I would emphasize this so much, because the traffic-related power is actually quite small... but that's a phrasing detail.
https://github.com/cpignata/powerbench/blob/ada3322da647a2b83ab4f3a97c6b1d12ada550f7/draft-cprjgf-bmwg-powerbench.xml#L185-L186
Actually, pretty much everything matters. The real challenge is to have a clean way of specifying and replaying the traffic trace you want to use. PCAPs don't scale well to 32*100G interfaces :-/
https://github.com/cpignata/powerbench/blob/ada3322da647a2b83ab4f3a97c6b1d12ada550f7/draft-cprjgf-bmwg-powerbench.xml#L192
Sure, everything must be reported. To me though, the important thing I'd hope to get from the document is not so much what I must report on, but how this reporting should look like.
https://github.com/cpignata/powerbench/blob/ada3322da647a2b83ab4f3a97c6b1d12ada550f7/draft-cprjgf-bmwg-powerbench.xml#L248-L249
I don't understand the "should be proportional to its traffic load" part. We know this is not the case, and will never be. Or do you mean just the dynamic power?