International Association
for Cryptologic Research

As Succinct Non-interactive Arguments of Knowledge (SNARKs) gain traction for large-scale applications, distributed proof generation is a promising technique to horizontally scale up the performance. In such protocols, the workload to generate SNARK proofs is distributed among a set of workers, potentially with the help of a coordinator. Desiderata include linear worker time (in the size of their sub-tasks), low coordination overhead, low communication complexity, and accountability (the coordinator can identify malicious workers). State-of-the-art schemes, however, do not achieve these properties.

In this paper, we introduced $\mathsf{Cirrus}$, the first accountable distributed proof generation protocol with linear computation complexity for all parties. $\mathsf{Cirrus}$ is based on HyperPlonk (EUROCRYPT'23) and therefore supports a universal trusted setup. $\mathsf{Cirrus}$ is horizontally scalable: proving statements about a circuit of size $O(MT)$ takes $O(T)$ time with $M$ workers. The per-machine communication cost of $\mathsf{Cirrus}$ is low, which is only logarithmic in the size of each sub-circuit. $\mathsf{Cirrus}$ is also accountable, and the verification overhead of the coordinator is efficient. We further devised a load balancing technique to make the workload of the coordinator independent of the size of each sub-circuit.

We implemented an end-to-end prototype of $\mathsf{Cirrus}$ and evaluated its performance on modestly powerful machines. Our results confirm the horizontal scalability of $\mathsf{Cirrus}$, and the proof generation time for circuits with $2^{25}$ gates is roughly $40$s using $32$ $8$-core machines. We also compared $\mathsf{Cirrus}$ with Hekaton (CCS'24), and $\mathsf{Cirrus}$ is faster when proving PLONK-friendly circuits such as Pedersen hash.