Neuromorphic Computing Approach for Intelligent Resource Allocation in High-Performance Cloud Platforms

Abstract

Existing heuristic and deep reinforcement learning (DRL) schedulers are inadequate for providing sub-millisecond decision latency and energy efficiency requirements for cloud platforms with tens of thousands of diverse workloads running in parallel. This paper proposes an online learning strategy, Spike-Timing Dependent Plasticity (STDP), and the Neuromorphic Scheduling Model (NMS), a brain-inspired resource allocation model involving Leaky Integrate-and-Fire (LIF) spiking neural networks (SNNs) and hybrid hardware layer, which integrates Intel Loihi-2 neuromorphic chips with conventional CPU/GPU cluster orchestration. NMS encodes workload telemetry as spike-rate and temporal spike patterns, facilitating event-driven scheduling decisions with a mean latency of 1.4 ms with a queue-depth of 500 jobs, which is 73.6% lower than the best deep reinforcement learning baseline. The experimental results on a 2K node testbed show that NMS achieves a 26.3% increase in the mean resource utilisation compared to the default Kubernetes scheduling, 41.8% decrease in job completion time, and 68.4% decrease in energy consumption under the same workload compared to scheduling engines accelerated by GPUs. All five key metrics show statistical significance (all p < 0.001) and large effect sizes (η² ∈ [0.76, 0.91]). It takes 142 training epochs for the closest DRL baseline, while STDP weight convergence takes place in 87 training epochs. The NMS architecture is benchmarked against ISO/IEC 25010 quality criteria and tested for scalability using polynomial regression modelling up to 50, 150, 300, 450, 950, 1,500 and 2,000 nodes per cluster.