Grid services and platforms that are always on and always available are awesome for running cheap, highly-available services on the internet. That’s great, because it means the barrier to entry, and the cost to deliver those services is low making innovations by small teams with little to no cash much more possible. What it means though, is that we’re creating Datacenters with tons of capacity that’s being used during two peak periods throughout the day. After speaking with a few Datacenter architects are the Sun Open House, I realized that not only are todays DC designs supremely inefficient, but most of the stuff running in them isn’t being used up to 80% of the day. Cooling costs are extremely high, and the cost of running a grid to sustain maximum capacity + wiggle room is shocking. So in order to reduce those costs some companies like HP and Sun are actively looking at computers that will turn themselves off or into ‘low power’ mode when not being used. That’s cool, but I don’t think we have to wait for them to build that system. If a computer isn’t required, then a remote ‘green monitor’ can log in and gracefully shutdown unnecessary systems, and the magic known as ‘Wake-on-LAN‘ can be used to bring them back up again. If you could build that into your grid then you could instantly save energy, not only to run a system, but on cooling as well. Of course, this would have to be mapped to a prediction engine that would say ‘traditionally we’ve needed more capacity at 7:30AM Pacific, so boot up some systems’, but that would be a really fun project.
Let’s take it one step further.
If energy costs are really high then you want to use as little of it as possible. So how about grabbing compute cycles from a grid that’s in the peak power zone for Solar? It’s cheap to send a few hundred gigs of data to be crunched, but its expensive to work on it. If you could do that, then you could have your capacity for computing literally follow the Sun. Someone has to work out the math on what’s cheaper - moving electricity, or moving the cycles, but my bets on the cycles because I think a mile of fiber costs less than a mile of High-voltage lines.
DC A wants to send a data crunching job offsite. It’s going to check the International Grid Exchange (I just made this up) for cheap capacity somewhere. DC B’s rate is 5.28$/teraflop hour because it’s in Iceland (which has cheap bandwidth and geothermal energy - oh and its cold). DC A’s rate for the same service is $8.22/teraflop hour, so DC A exports that task to DC B, and when it’s done DC A gets the resulting data set - or a trickle of data comes back over the tasks compute time.
Now we have a market where the most efficient winner gets the utilization. True utility computing, true scaling on demand, no market hype or hand waving.
You could even buy Grid ‘Futures’; You want to make sure you get the capacity, or you could prospect and buy chunks of time for resale later on. Basically turning computing into a true commodity.
I hope that’s one way to make computing a lot more efficient.