[Paper Review] nOS-V: Co-Executing HPC Applications Using System-Wide Task Scheduling

David Álvarez, Kevin Sala+1 more|arXiv (Cornell University)|Apr 22, 2022

Distributed and Parallel Computing Systems4 citations

TL;DR

This paper introduces nOS-V, a lightweight tasking library enabling system-wide, fine-grained co-execution of HPC applications on shared nodes by unifying task scheduling across multiple applications. By replacing individual runtimes with a single global scheduler, nOS-V reduces idle resources and improves efficiency, achieving a median 1.25x speedup for three-way co-execution and 1.17x for pairwise combinations compared to exclusive execution.

ABSTRACT

Future Exascale systems will feature massive parallelism, many-core processors and heterogeneous architectures. In this scenario, it is increasingly difficult for HPC applications to fully and efficiently utilize the resources in system nodes. Moreover, the increased parallelism exacerbates the effects of existing inefficiencies in current applications. Research has shown that co-scheduling applications to share system nodes instead of executing each application exclusively can increase resource utilization and efficiency. Nevertheless, the current oversubscription and co-location techniques to share nodes have several drawbacks which limit their applicability and make them very application-dependent. This paper presents co-execution through system-wide scheduling. Co-execution is a novel fine-grained technique to execute multiple HPC applications simultaneously on the same node, outperforming current state-of-the-art approaches. We implement this technique in nOS-V, a lightweight tasking library that supports co-execution through system-wide task scheduling. Moreover, nOS-V can be easily integrated with existing programming models, requiring no changes to user applications. We showcase how co-execution with nOS-V significantly reduces schedule makespan for several applications on single node and distributed environments, outperforming prior node-sharing techniques.

Motivation & Objective

• To address the inefficiency of exclusive node execution in modern HPC systems, where applications often underutilize resources due to serial phases, load imbalance, or lack of parallelism.
• To overcome limitations of existing co-location and oversubscription techniques that suffer from interference, lack of global visibility, or application-specific constraints.
• To design a portable, lightweight runtime system that enables fine-grained, system-wide scheduling of HPC workloads without requiring application modifications.
• To evaluate the performance of co-execution across diverse workloads, including single-node and distributed configurations with varying NUMA policies.

Proposed method

• nOS-V implements a unified tasking runtime that manages all tasks from co-scheduled HPC applications globally, replacing per-application runtimes.
• It leverages standard inter-process communication (IPC) mechanisms and requires no kernel modifications, ensuring portability and ease of integration.
• The system maintains a global task queue and applies scheduling policies based on process priority or data locality to optimize performance.
• nOS-V supports co-execution of multiple applications—including hybrid MPI and task-based programs—without requiring changes to user code or binaries.
• It enables dynamic load balancing by scheduling tasks across available cores based on real-time availability and locality, avoiding interference from OS preemptions.
• The approach is integrated with the Nanos6 runtime and OmpSs-2 programming model, enabling evaluation on real HPC workloads.

Experimental results

Research questions

• RQ1 Can system-wide task scheduling significantly improve node-level resource utilization in HPC workloads with varying degrees of parallelism?
• RQ2 How does co-execution via a unified runtime compare to static co-location and dynamic oversubscription in terms of performance and scalability?
• RQ3 To what extent can co-execution reduce schedule makespan across diverse HPC applications, including those with NUMA-aware memory access patterns?
• RQ4 Can nOS-V support complex workloads such as three-way co-execution and distributed applications without performance degradation?

Key findings

• nOS-V achieves a median 1.17x speedup over exclusive execution for pairwise co-execution of HPC applications on single nodes.
• For three-way co-execution, the median speedup increases to 1.25x, demonstrating improved scalability and resource utilization.
• The system reduces remote NUMA accesses by assigning tasks to cores on the same socket, improving data locality and reducing latency.
• nOS-V outperforms existing co-location techniques like DLB, Callisto, and AMCilk by providing a global view of all tasks and avoiding interference from per-application runtimes.
• The integration with OmpSs-2 and Nanos6 enables transparent co-execution without modifying user applications or binaries.
• The approach is effective across diverse workloads, including those with irregular parallelism and heterogeneous memory access patterns.