llnl/Umpire

An application-focused API for memory management on NUMA & GPU architectures

What it solves

Umpire addresses the challenge of managing memory across complex hardware architectures, specifically machines that utilize multiple memory devices such as GPUs and NUMA (Non-Uniform Memory Access) systems.

How it works

It functions as a resource management library that enables the discovery, provisioning, and management of memory resources across these diverse hardware devices.

Who it’s for

Developers working on high-performance computing applications that need to efficiently coordinate memory across hierarchical and heterogeneous memory devices.

Highlights

  • Resource management for multiple memory devices.
  • Support for NUMA and GPU memory.
  • Discovery and provisioning of memory resources.

Related

  • Project
    rapidsai/rmm

    NVIDIA RMM (RAPIDS Memory Manager) is a C++/Python library that provides flexible, high‑performance GPU memory allocation. It offers multiple allocation strategies (plain, pooled, managed, binning), stream‑ordered allocation, per‑device default resources, and easy CMake/Python integration, helping GPU‑centric workloads avoid the overhead of repeated `cudaMalloc`/`cudaFree`.

  • Project
    nv-legate/cupynumeric

    A high-performance array computing library that implements the NumPy API on top of the Legate framework to scale NumPy workflows across multiple GPUs and nodes.

  • Project
    anderskm/gputil

    A Python module that uses nvidia-smi to monitor NVIDIA GPU status and programmatically select available GPUs based on memory and load.

  • Project
    Traverse-Research/gpu-allocator

    A Rust-native GPU memory allocator that provides a unified interface for managing memory across Vulkan, DirectX 12, and Metal.

  • Project
    Project-HAMi/HAMi-core

    An in-container GPU resource controller that intercepts CUDA calls to enforce per-container memory and compute utilization limits without modifying applications or drivers.