/SALLOC

Smart allocators for SIMT architectures

Primary LanguageC++Apache License 2.0Apache-2.0

SALLOC: Smart allocators for SIMT architectures

SALLOC is a thread-safe arena allocator (memory manager) in CUDA.

SALLOC has been developed as part of the project "Smart Data Structures in CUDA" with CERN-HSF for Google Summer of Code 2017.

Features

  • Thread-safe arena allocator for GPUs in CUDA.
  • Supports thread-safe vector container in CUDA.
  • Support for multiple vectors on the arena. These vectors are shared across threads.
  • The vector container supports the following operations:
    1. push_back()
    2. pop_back()
    3. getIndex()
    4. vecSize()

Documentation

A detailed documentation on SALLOC can be found in SALLOC_description.txt.

Requirements

Linux

  • CUDA version >= 8.0
  • gcc version >= 5.3.0

Getting Started

Getting the code

  • Clone the code into a directory called SALLOC

    git clone --branch=master https://github.com/ssomesh/SALLOC

  • The source for arena allocator SALLOC is present in the file salloc.h

  • No installation or build is required.

Using SALLOC in your code

  • Include "salloc.h" (with the proper path) in the CUDA code and use.
  • driver.cu contains a sample program that uses SALLOC.
    • Compile driver.cu and execute 
      nvcc driver.cu -std=c++11 -o driver

./driver
  • SALLOC has been tested on NVIDIA Pascal (GeForce GTX 1080) GPU.

TODOs

  • Support resize() operation on vector container allocated on the arena.
  • Support reclaiming of free chunks in the arena to better utilize the space on the arena.

Developers

  • Felice Pantaleo (felice.pantaleo (at) cern.ch)
  • Somesh Singh (somesh.singh1992 (at) gmail.com)