Petascale Urban Computing
by Vinay Gupta • March 31, 2013 • Everything Else • 4 Comments
Simple concept (!) I wanted to give a URL to.
- One thousand or more nodes in a mesh or other network
- Each node has:
- One terabyte or more of disk
- One gigabyte or more of ram
- On gigahertz or more of processor
- Nodes are scattered through a contiguous geography like a city
Conceptually, Petascale Urban Computing weaves a supercomputer through a city. A petabyte is enough to store most of what a community is interested in, and the combined processing power is plenty to be getting on with.
How do you program a petascale urban computer? One possibility is JavaScript – nodes can relatively safely run JS programs for other nodes, and they can’t touch local disk. Search, transcoding etc. could all be done with such mechanisms with very thing APIs exposed by REST or JSON calls to helper apps on local web servers (i.e. your mobile JS agents call local services through http://127.0.0.1/do_stuff_for_me type calls.)
The core vision is that your PUC router is a direct connection to an urban supercomputer that completely penetrates your world, continually crunching data so that whatever you need is right there when you want it.
At night, when you sleep, it thinks on your behalf.
I have a http://www.kickstarter.com/projects/adapteva/parallella-a-supercomputer-for-everyone on order. Should be able to run node.js or the Erlang OTP on it.
Fun to imagine one of these in every house or flat.
And for a weird little bit of contrast, and perhaps a lesson in the top-down v. bottom-up arena: http://arstechnica.com/information-technology/2013/04/worlds-first-petascale-supercomputer-will-be-shredded-to-bits/
Pingback: People’s Cloud architecture and futures in 1000 words | The Bucky-Gandhi Design Institution
Vinay, you mentioned at EMW that Whitespace might serve a useful purpose here.
Although I think multiple thousands of nodes within one cell is well suited to whitespace, I’m not sure the throughput (20Mbps) is enough for data-intensive activities & maybe meshed short-range might be suitable.
What I think could be interesting is a software-defined radio (SDR) device with a metamaterial antenna requesting short-lease spectrum using a dynamic spectrum access database. A device with these capabilities could reconfigure to any available wavelength using the correct radio characteristics and then search the area for usable frequencies, maybe using a bootstrap channel to the database.
The PUC devices would fire up clusters which could settle on a communication model and frequency according to their environment. Perhaps the clusters would work as their own parallel processor in a geographical area, meaning you move into each coverage area like you imagine a mobile cell or wifi hotspot now.
When do we start?