We address fundamental problems in distributed systems using protocols, data structures and algorithms inspired by Quantum Information Theory.
Our market is next generation platforms for secure, reliable, cutting edge, distributed computing. We provide 5G network slices with a fundamentally more secure graph confinement architecture.
Initial use-cases include Digital Twins, Multiplayer Games and Web 3.0.
The Problem
Reliability is a perennial challenge in distributed systems today. Applications typically reside on different computers that communicate over fallible networks. Packets are routinely dropped, reordered, duplicated or delayed. Conventional switched networks rely on ad hoc methods such as timeouts and retries to maintain liveness. This makes exactly-once semantics impossible, and precipitates retry storms, leading to unbounded tail latency, and significant rates of transaction failure. 80% of failures have a catastrophic impact, with data loss being the most common (27%). 90% of the failures are silent; the rest produce warnings that are unclear. 21% of the failures lead to permanent damage to the system, and this damage persists even after the network partition heals.
The Solution
Daedaelus is a company founded by Paul Borrill who spent decades thinking about, and working on solutions to network reliability problems, and who put together a team of experts dedicated to solving them. The result is a completely new kind of software that can radically transform workplace productivity. We expect to achieve an order of magnitude improvement in: reliability, deployment time, reductions in administrative effort. And at least 50% improvement in energy utilization for the same workloads.
How Does it Work?
Central to our solution is a reversible Distributed Atomic Ethernet protocol in an FPGA substructure (between SmartNICs in one or more racks), using the Stop and Wait protocol to maintain flow control, Colored Petri Nets for conserved quantities, and the Spekkens Toy Model to manage epistemically restricted consistency for distributed metadata. Our rack-scale microdatacenters are scale-independent; we begin with 9, 25, or 100 nodes, but there are no limits except cost (CapEx and OpEx).
Daedaelus’ initial product uses as a node a server, with a PCIe slot, and an FPGA card in it, with 6 to 8 cables connected to its neighbors in a plane. Using an FPGA means that some of the hardware logic can be designated as a complete working switch without relying on software. This lowers power consumption and latency.We use a mesh network of chiplet servers: small-scale hexagonal stacked circuits with no network switches. The mesh network of chiplet servers enables true peer to peer communication over Compute Express Link, and provides a significant cost reduction relative to best practices.
Why Does it Work?
The Daedaelus solution reduces latency in ways that conventional networks cannot. First, it makes direct connections. Second, it works by way of multicast consensus, in parallel over 8 ports instead of serial over 1. Third, it truncates tail latency; the protocol knows it failed or succeeded (without heartbeats or timeouts). There are no dropped packets. If a link fails, we re-route, locally, around the failure. If a packet doesn’t reach its destination, we know immediately. The protocol ensures that both ends have the same facts about whether that packet was delivered or not.
For More Information
You can find more detailed technical specs, as well as slides for a presentation on the company that was recently given at the Chiplets Summit Here.