Deterministic parallelism is a major building block for distributed and fault-tolerant systems, offering substantial performance benefits with its usecases extending to debugging and testing as well. By studying recent Deterministically Parallel Systems (DPS), we identify several architectural limitations that hinder their performance and efficiency. As a result, existing DPS cannot keep up with the high-throughput and 𝜇𝑠-scale distributed systems that are deployed in modern datacenters today.

We introduce DORADD, a high-performance deterministically parallel runtime, which is tailored for modern datacenter applications. DORADD takes a proactive approach to enforce deterministic parallel execution entirely through scheduling, thus allowing requests to run to completion in a synchronization-free manner. It dynamically captures inter-request dependencies and schedules requests on the available CPU resources in a work-conserving manner. It proposes a novel mechanism, i.e., core pipelining, to scale the single-dispatcher performance. We build an in-memory database with DORADD and compare it with Caracal, the current state-of-the-art deterministic database, using the YCSB and TPC-C benchmarks. Our evaluation shows up to 2.5x better throughput and more than 150x and 300x better tail latency in non-contended and contended cases. We also compare DORADD with Caladan, the state-of-the-art non-deterministic scheduler, showing that determinism does not induce any throughput overhead under latency SLAs.