What if you could harness all the unused compute capacity in your company’s servers, workstations, IoT devices and even phones to create a supercomputer without the expense of buying one?
That’s the idea behind Computes, which recently launched its beta.
Anything with a CPU or GPU can join a private, secure mesh network that together can create what founder Chris Mattieu calls a mesh supercomputer. The result is enough compute power to perform serial-based machine learning algorithms as well as massively parallel supercomputing tasks even on edge networks where machine data is being generated. And without the need to buy any more infrastructure.
The University of Wisconsin, for one, is using the Computes technology in its Parkinson’s research. Carrie Rountrey, an assistant professor in the Department of Communications Sciences and Disorders, has developed an algorithm that looks at speech patterns in voicemail messages to predict early onset of the disease.
The university has 20,000 workstations it wants to leverage to support research.
Computing at the Edge
Mattieu admits he’s obsessed with the Skynet artificial intelligence system in the Terminator movies, and originally named his previous company that. Later renamed Octoblu, it was acquired by Citrix in December 2014.
Octoblu was a system for drones to talk to each other. It was designed to run on a single network or mesh of IoT networks that share a common API or communications protocol enabling devices to discover, query, and message other devices on the network.
Citrix made that technology open source but decided to sunset development of it internally.
The early idea for Computes was built as a centralized cloud supercomputer, where the cloud distributed work out to the nodes, but “I quickly realized all these lines between cloud and edge really shouldn’t exist,” Mattieu said. “Everyone’s trying to move to the edge, like self-driving cars don’t have time to make a request to the cloud then do some learning and come back with an answer. Everything has to happen where the computing is originating.”
The network is decentralized and distributed, based on what Mattieu calls “blockless technology.”
With blockchain technology, you write data to a “block,” where you cryptographically hash, mine and sign the block to it to the database. Every computer on the chain has a copy of the entire database, which means IoT devices are too small to even be part of this, and it’s slow, he explained.
Bitcoin, the most famous one, has a throughput of only three transactions per second while Ethereum is up to about 13 per second. By contrast, Visa’s network handles 24,000 transactions per second.
Computes instead uses technology similar to IOTA, which uses a directed acyclic graph (DAG) or Merkle tree for storing transactions. It’s used on GitHub, in which you can track each code commit and trace it back to the branch on the tree from which it originated and find associated time stamps for each commit.
It doesn’t require proof of work for mining like on the blockchain network and doesn’t require each device to maintain an entire database, making it much faster.
To create the self-organizing and self-healing network, each device needs a 5MB software “nanocore.”
Everything is encrypted, both in transit and at rest. Only that company’s nanocores know how to swarm and organize themselves on that private mesh network, he said. It uses IPFS (Interplanetary File System), a peer-to-peer system for sharing files that also has a pub/sub messaging protocol. It allows all the nodes to message each other and talk about what needs to be done on which nanocore. Then as those nanocores request the work, there are hashes for IPFS files for input, and those cores return an IPFS hash for where the output file is in this mesh network.
For now, the Phoenix, Ariz.-based company is focused on large enterprises that need massive compute power, such as for scientific research.
Later, there might be opportunities for the general public to make a few bucks by selling their excess capacity, he said.
The National Parkinson’s Foundation also is looking at the prospect that it might be able to combine computes among patients and supporters that might want to donate their excess CPU and GPU resources to research projects, he said.
Feature image via Pixabay.
