MC-SPARK: a Policy-Constrained Genetic Algorithm for Multi-Cloud Workload Placement

Abstract

Multi-cloud adoption expands pricing, performance, geographic, and resilience options, but makes workload placement combinatorial and policy dependent. This paper presents MC-SPARK, a policy-constrained genetic algorithm that maps workloads to candidate services while jointly considering operating cost, latency risk, a carbon proxy, and migration effort. Hard constraints enforce provider eligibility, capacity, availability, and regional requirements. A controlled synthetic study uses 60 heterogeneous workloads and 24 normalized services across three providers and three regions. Across 15 independent runs, MC-SPARK reduces mean monthly cost by 49.2%, latency penalty by 69.1%, and carbon proxy by 50.1% relative to current placement, with zero hard-policy violations. A sensitivity study, 95% confidence intervals, hypothesis tests, comparison with NSGA-II, and scalability measurements are included. The results demonstrate algorithmic behavior rather than provider-specific price superiority.