Permissioned ledgers allow users to execute transactions against a data store, and retain proof of their execution in a replicated ledger. Each replica verifies the transactions’ execution and ensures that, in perpetuity, a committed transaction cannot be removed from the ledger. Unfortunately, this is not guaranteed by today’s permissioned ledgers, which can be re-written if an arbitrary number of replicas collude. In addition, the transaction throughput of permissioned ledgers is low, hampering real-world deployments, by not taking advantage of multi-core CPUs and hardware accelerators.
This thesis explores how permissioned ledgers and their consensus protocols can be made auditable in perpetuity; even when all replicas collude and re-write the ledger. It also addresses how Byzantine consensus protocols can be changed to increase the execution throughput of complex transactions. This thesis makes the following contributions:
1. Always auditable Byzantine consensus protocols. We present a permissioned ledger system that can assign blame to individual replicas regardless of how many of them misbehave. This is achieved by signing and storing consensus protocol messages in the ledger and providing clients with signed, universally-verifiable receipts.
2. Performant transaction execution with hardware accelerators. Next, we describe a cloud-based ML inference service that provides strong integrity guarantees, while staying compatible with current inference APIs. We change the Byzantine consensus protocol to execute machine learning (ML) inference computation on GPUs to optimize throughput and latency of ML inference computation.
3. Parallel transactions execution on multi-core CPUs. Finally, we introduce a permissioned ledger that executes transactions, in parallel, on multi-core CPUs. We separate the execution of transactions between the primary and secondary replicas. The primary replica executes transactions on multiple CPU cores and creates a dependency graph of the transactions that the backup replicas utilize to execute transactions in parallel.