Heterogeneous Distributed Parallel Processing via gRPC with Zero-Configuration UDP Worker Discovery

Abstract

Modern computational workloads, including large-scale Monte Carlo simulations and distributed machine learning training, demand parallelism that exceeds the capacity of a single machine. Existing distributed computing frameworks like Apache Spark and Berkeley Open Infrastructure for Network Computing (BOINC) require significant infrastructure overhead, static network configuration, and homogeneous runtime environments, limiting their accessibility in ad-hoc settings such as student laboratories or home networks. This paper proposes a zero-configuration heterogeneous distributed parallel processing framework based on a three-tier master-worker architecture, wherein worker nodes autonomously discover and register with a central orchestrator via a User Datagram Protocol (UDP) broadcast beaconing protocol, eliminating all manual network configuration. Task payloads comprising executable logic, data shards, and model state are dispatched to worker nodes via Google Remote Procedure Call (gRPC) over HTTP/2 using Protocol Buffers (Protobuf) binary serialization, achieving a significant payload size reduction compared to equivalent JSON representations. Worker liveness is maintained through an asynchronous heartbeat mechanism with automatic shard re-assignment upon node failure, providing fault tolerance with zero manual intervention. The framework is validated through a 200,000,000-trial Geometric Brownian Motion (GBM) Monte Carlo simulation distributed across a three-node heterogeneous cluster, specifically chosen to test extreme cross-platform heterogeneity. A superlinear speedup of 4.29x was observed over single-node execution with a measured parallel efficiency of 143%, attributable to the aggregate multi-core advantage of the heterogeneous worker pool. The proposed system demonstrates that enterprise-grade distributed computing capabilities can be realized without dedicated infrastructure, offering a practical and extensible platform for parallel scientific computation and distributed machine learning.