Recently I was interacting with some data center designers and architects and some other vendors when the topic of power draw came up. This is always a fun one, and an area where a lot of 'specsmanship' is being played nowadays as people try to find creative ways to showcase their power efficiency, or hide/mask power draw that is less than optimum. Needless to say it makes for a fun conversation or two.The question that was asked though was, "How much power does your product consume?" This seems, at first glance, to be a relatively straightforward question, most likely answered in some amount of Watts consumed over some given interval, usually an hour. So a very blunt answer would be something to the effect of, "My product consumes 3.6kWh, or 3,600 Watt hours." (This would be like leaving two of those really powerful industrial strength hair-dryers they sell at Wal-Greens on for an hour on high heat and high fan) But I didn't hear a very straightforward answer - I heard the following questions/statements:I've heard of 'Nominal Power' before, but usually in the field of automotive audio, where stereo manufacturers like to talk about their 1500W subwoofers and such, but then list the 'Nominal Power' which is far less output and is more what the product can sustain when not 'bursting' to a capacitor draining and ear-shattering peak. This seemed directionally accurate, but still off a bit so I wondered what network infrastructure manufacturers must have meant... so I dug a bit deeper, and turned to Dictionary.com and Wikipedia, the sources of all truth in this millennium.The abbreviated version of their answers that came across as marginally relevant are:–adjectiveThese seem to fit, although again a bit of a stretch, excepting maybe for #1. See when I asked the vendor what they meant I was told, "This is our power draw under average load and in a 20C ambient air environment with controlled humidity." Basically, about as best-case a number as you can hope for.See, there are a lot of variables to power draw. Let's take switches, it is a product I am familiar with, lots of people have them, and since this is NetworkWorld I don't have to explain what they are, or why they matter- which is rather nice :)The main design-level factors that cause power consumption on the switch are:
Nominal Power, right?
Well, how about power per port?
Under what load and environmental conditions?
1. being such in name only; so-called; putative: a nominal treaty; the nominal head of the country.
2. (of a price, consideration, etc.) named as a mere matter of form, being trifling in comparison with the actual value; minimal.
7. (of money, income, or the like) measured in an amount rather than in real value: Nominal wages have risen 50 percent, but real wages are down because of inflation.
8. Aerospace . performing or achieved within expected, acceptable limits; normal and satisfactory: The mission was nominal throughout.
- The number of ASICs, memory assemblies, processors, FPGAs, and NPUs necessary to support the control and data plane
- The speed and geometry of these ASICs
- The lasers (optics) and PHYs (copper) for transmission
- The fans to move the air across the heat sinks of the board-level components mentioned above
- Conversion loss from AC to DC and DC to DC as power is distributed to a series of components that operate at different voltages
The main environment and variable factors that modulate power consumption from minimum to maximum are:
- The ambient air temperature
- The load on the device
- The types of optics or physical media devices plugged into the ports on the device
What all of this is really leading to though is a simple question -- Why are we not just making planning decisions on max numbers?Sure for some things where you can feasibly oversubscribe a service it makes sense to enable oversubscription - (network traffic in the campus is a classic and ideal scenario where there is almost no need for max capacity planning, noise is probably another where the worst case scenario is either you wear headphones or feel like you were at a Spinal Tap concert for a few hours.) But we are talking about power, where not provisioning enough power creates safety hazards, and cooling, where not provisioning enough cooling capacity puts the entire data center at risk of critical component outages because of overheating and auto-shut down. I mean, do we really want to have to explain to anyone that we caused a widespread outage of a data center because we didn't provision enough cooling because we planned for 'nominal' and didn't anticipate 'max' ever happening?So this leaves me with the following, and final thoughts: Vendors need to publish max power draw numbers, at a system level, for all products. Sure we can put out 'nominal' or average power draw numbers that look more appealing in side-by-side data sheet bake-offs. But when it comes down to it, and its safety or availability on the line, and OPEX dollars off the bottom-line every quarter. We should normalize all the network performance variables to meet whatever our oversubscription, buffering, route table size, or ACL depth metrics are for our designs, and then measure the max power draw to support a fully loaded system with the optics we need at 100% load at 'hot summer day, failed HVAC' type of cooling levels, and then determine the best products for our networks.Yes, its getting to be a more complex world... but its still 'nominally' fun!dg
Most network equipment worth deploying in an enterprise has some form of intelligent power management. The smarter systems use a combination of testing and modeling to determine the optimum fan rotational velocity to cool the system so components do not overheat and fail at a given load factor and ambient air temperature setting. The slightly less intelligent ones look at heat sensors on most components and spin up the fans faster when they get warm, slow them down when not so warm. The slower the fans rotate the less power it takes to cool the system. The lower the ambient temperature the slower the fans can spin, but in many cases, especially in the warmer months, it takes more power to cool the air to an ambient temperature that appreciably impacts the fan speeds than it would take to just run the data center warmer and spin the fans faster.