Cerberus Consensus: Peer Reviewed | The Radix Blog | Radix DLT
RDX Works is pleased to announce that the academic paper evaluating Cerberus, the novel cross-shard consensus protocol powering Radix, has now been accepted to JSys, the Journal of Systems Research, after completing a rigorous peer review process. The peer review involved independent experts verifying the proof, theory, and soundness of Cerberus.
The paper provides analysis of different solutions to multi-shard BFT consensus to demonstrate that Cerberus is the most performant and efficient consensus protocol that supports atomic updates across shards. Other notable consensus protocols evaluated included Chainspace, which was acquired for Facebook’s Libra DLT solution, AHL, Sharper, and Ring BFT.
RDX would like to thank Professor Mohammad Sadoghi and the University of California Davis for their support in the analysis and testing of the various protocols contained within the paper and during the peer review process.
Professor Sadoghi said:
In partnership with Radix over the past four years, we have been exploring the frontier of resilient consensus, with patience and perseverance overcoming the skepticism and technical barriers in this uncharted territory. Along the way we have educated the public inside & outside the classroom, while focusing on our joint vision to build sustainable global-scale blockchain platforms and above all cherishing human co-operation and collaboration on this fulfilling journey with friendship.
Radix now joins a small family of public ledgers that have had their consensus protocol tested to the highest academic standards.
You can read the paper here, or read our explanation of the paper below.
Explanation of the Paper
The majority of smart contract platforms today, such as Ethereum, are “fully replicated”, which limits the performance of the overall system to the speed of an individual node, ruling out scalability.
To address this, a number of “multi-shard” consensus protocols have emerged. These process transactions across shards, providing the ability to scale distributed ledgers linearly with the number of shards. Some of these state of the art consensus protocols include AHL, ByShard, Caper, Chainspace, RingBFT, and SharPer.
Comparison of key attributes for the different consensus protocols
This paper, Cerberus: Minimalistic Multi-shard Byzantine-resilient Transaction Processing (June 2023), introduces a new problem to this field of cutting edge academia: how do you minimize the coordination cost for multi-shard consensus protocols? How do you make the protocol as efficient and performant as possible?
To do this, the paper presents Cerberus as a new family of consensus protocol, with three variants: Core, Optimistic, and Resilient; each fine tuned for different use cases.
The key innovations Cerberus introduces is the means in which a disjointed validator sets communicate with each other efficiently with low complexity to arrive at an agreement, and the ability to process many UTXO-like objects concurrently within the same shard without explicit locking or blocking of the shard.
The paper then took Cerberus and its peer consensus protocols through a variety of rigorous tests.
This testing was performed in a permissioned environment where each replica (node) had a bandwidth of 1Gbit/s and message delay of 15ms. Then, various scenarios were tested that changed the number of replicas per shard, objects per transaction, and number of shards. Scenarios where there were malicious nodes were also included, testing how the various protocols would perform under adverse conditions.
Criteria to compare the protocols included how efficient they were at conducting consensus as well as how they fared under a variety of failure modes.
What linear scalability looks like, with Cerberus processing over 1m transactions per second (TPS) in some scenarios
As can be seen above, all three flavors of Cerberus clearly came out on top under all scenarios.
The linear scalability of Cerberus is clearly evident, and tests showed that Cerberus comfortably exceeded 1 million transactions per second under multiple scenarios. With linear scalability, there is no reason why 10 million or 100 million transactions per second couldn’t be achieved so long as enough shards are added.
The conclusion from all these tests? Cerberus is the most efficient consensus protocol, has the best throughput, and the lowest latency when compared to other state of the art multi-shard consensus protocols.
We look forward to the implementation of sharded Cerberus as part of the Xi’an Mainnet Upgrade (the next mainnet upgrade after the Babylon Mainnet upgrade).
Some Questions You May Have
What’s changed since 2020?
The consensus protocol described in Cerberus – A Parallelized BFT Consensus Protocol for Radix (March 2020) and Cerberus: Minimalistic Multi-shard Byzantine-resilient Transaction Processing (August 2020) has hardly changed.
The main difference is that the paper has been through a lot more academic rigour, with terms being more precise and described in line with academic standards:
- The detail regarding the bounds and assumptions the protocol operates within have been improved and are now more clear.
- There is more explanation of how Cerberus achieves consensus and the security model.
- More edge and corner cases have been explored and how Cerberus resolves them have been described.
What assumptions does the paper make?
To get the full list, you would have to read the paper.
But as headlines:
- The consensus protocols were tested in a permissioned environment.
- Adversaries are strong, all knowing, but not all influencing.
- Under a condition of partial or full asynchrony, the network will eventually revert to a state of synchrony
- Messages will eventually arrive at their destination, but may require retransmission.
- Sybil resistance is not in the scope of the paper as it is a different problem domain to consensus. Addressing Sybil is not required to prove the soundness and robustness of a consensus protocol and would complicate the paper and blur its focus. Furthermore it is common not to do so in academic papers which are focussed on the problem of consensus.
What did the reviewers look for?
Tightening of definitions, taking Cerberus apart from a theoretical perspective, and whether the testing methodology and conclusions were sound.
How were the other consensus protocols selected?
A very simplified version of the criteria for selection was for consensus protocols that could be capable of processing billions of transactions a day, and could support atomic updates across shards.
To learn more about Cerberus and how it fits into the rest of Radix, see The Radix Full Stack and How Radix Engine is Designed to Scale dApps. To really get into the detail and learn how Cerberus works with pictures, be sure to check out the Cerberus Infographic Series.