« Stuff The Internet Says On Scalability For January 8th, 2016 | Main | Sponsored Post: Netflix,, Redis Labs,, SignalFx, InMemory.Net, VividCortex, MemSQL, Scalyr, AiScaler, AppDynamics, ManageEngine, Site24x7 »

Let's Donate Our Organs and Unused Cloud Cycles to Science

There’s a long history of donating spare compute cycles for worthy causes. Most of those efforts were started in the Desktop Age. Now, in the Cloud Age, how can we donate spare compute capacity? How about through a private spot market?

There are cycles to spare. Public Cloud Usage trends:

  • Instances are underutilized with average utilization rates between 8-9%

  • 24% of instance reservations are unused

Maybe all that CapEx sunk into Reserved Instances can be put to some use? Maybe over provisioned instances could be added to the resource pool as well? That’s a lot of power Captain. How could it be put to good use?

There is a need to crunch data. For science. Here’s a great example as described in This is how you count all the trees on Earth. The idea is simple: from satellite pictures count the number of trees. It’s an embarrassingly parallel problem, perfect for the cloud. NASA had a problem. Their cloud is embarrassingly tiny. 400 hypervisors shared amongst many projects. Analysing all the data would would take 10 months. An unthinkable amount of time in this Real-time Age. So they used the spot market on AWS.

The upshot? The test run cost a measly $80, which means that NASA can process data collected for an entire UTM zone for just $250. The cost for all 11 UTM zones in sub-Sarahan Africa and the use of all four satellites comes in at just $11,000.

“We have turned what was a $200,000 job into a $10,000 job and we went from 100 days to 10 days [to complete],” said Hoot. “That is something scientists can build easily into their budget proposals.”

That last quote, That is something scientists can build easily into their budget proposals, stuck in my craw.

Imagine how much science could get done if you didn’t have the budget proposal process slowing down the future? Especially when we know there are so many free cycles available that are already attached to well supported data processing pipelines. How could those cycles be freed up to serve a higher purpose?

Netflix shows the way with their internal spot market. Netflix has so many cloud resources at their disposal, a pool of 12,000 unused reserved instances at peak times, that they created their own internal spot market to drive better utilization. The whole beautiful setup is described Creating Your Own EC2 Spot Market, Creating Your Own EC2 Spot Market -- Part 2, and in High Quality Video Encoding at Scale.

The win: By leveraging the internal spot market Netflix measured the equivalent of a 210% increase in encoding capacity.

Netflix has a long and glorious history of sharing and open sourcing their tools. It seems likely when they perfect their spot market infrastructure it could be made generally available.

Perhaps the Netflix spot market could be extended so unused resources across the Clouds could advertise themselves for automatic integration into a spot market usable by scientists to crunch data and solve important world problems.

Perhaps donated cycles could even be charitable contributions that could help offset the cost of the resource? My wife is a tax accountant and she says this is actually true, under the right circumstances.

This kind of idea has a long history with me. When AWS first started, I like a lot of people wondered, how can I make money off this gold rush? That’s before we knew Amazon was going to make most of the tools to sell to the miners themselves. The idea of exploiting underutilized resources fascinated me for some reason. That is, after all, what VMs do for physical hardware, exploit the underutilized resources of powerful machines. And it is in some ways the idea behind our modern economy. Yet even today software architectures aren’t such that we reach anything close to full utilization of our hardware resources. What I wanted to do was create a memcached system that allowed developers to sell their unused memory capacity (and later CPU, network, storage) to other developers as cheap dynamic pools of memcached storage. Get your cache dirt cheap and developers could make some money back on underused resources. A very similar idea to the spot market notion. But without homomorphic encryption the security issues were daunting, even assuming Amazon would allow it. With the advent of the Container Age sharing a VM is now way more secure and Amazon shouldn’t have a problem with the idea if it’s for science. I hope.

Reader Comments (4)

I am trying to work this idea out.

As an EC2 customer I can select an option to donate my unused resources to "science"? This would effectively put these unused resources into a "science" pool that EC2 would give away to worthwhile scientific projects.

I like the idea, IMHO what this idea is missing is some sort of catchy mission statement, something that yields big PR value for Amazon as well as the customer. Something like: BigCompanyX donates its spare cycles to ImportantCauseY via Amazon's #ScientistsComputeForFree initiative to save lives in ThirdWorldCountryZ. It would be a PR write-off by BigCompanyX and Amazon would be the facilitator (win-win)

January 6, 2016 | Unregistered CommenterRussell Sullivan

I wasn't thinking of it as an EC2 feature.. More of a service that was run voluntarily, using if possible, the private spot instance software Netflix is developing. It would be cool if AWS would do such a thing, but that seems unlikely.

January 6, 2016 | Registered CommenterTodd Hoff

That's a good way to drive the pricing up. Currently, those "unused" cycles theoretically allow machines to be oversold. Even without that, they're consuming far less power than they otherwise would be. Start maxing them out and you'll watch your pricing start to rise as well - along with an increased carbon footprint, cooling issues, &c.

January 13, 2016 | Unregistered CommenterRichard

For the record, I think a system that forms an untrusted memcache cluster using spare cycles from strangers' EC2 instances is doable and has (some) value.

Richard is right, such a system could drive up prices or Amazon could just outlaw it. But if such a system existed, didnt expose your data, could provide reasonable SLAs, and was significantly cheaper than any other memcache option, why not use it.

Systems like this are tough to build in practice as the server is not trusted, so the client has to pre-encrypt all keys and values and pack them with various metadata that can be used on retrieval to validate correctness and freshness of the data.

Additionally there are the dependability & capacity planning issues, imagine using this system smoothly for months and then cyber monday rolls around and there is a widespread resource drought.

So a system like this is doable in theory, but it would not succeed in practice

January 16, 2016 | Unregistered CommenterRussell Sullivan

PostPost a New Comment

Enter your information below to add a new comment.
Author Email (optional):
Author URL (optional):
Some HTML allowed: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <code> <em> <i> <strike> <strong>