depaulmillz / lslab

Implementation of the L-slab hash map introduced in KVCG.

lslab

Designed in "KVCG: A Heterogeneous Key-Value Store for Skewed Workloads" by dePaul Miller, Jacob Nelson, Ahmed Hassan, and Roberto Palmieri.

Hardware

Runs on Volta, sm_70, or greater.

Building

Requires CMake >= 3.18, Conan, and a CUDA version in 11.0 to 12.1.

We also require UnifiedMemoryGroupAllocation built through conan.

Get CMake from kitware and Conan from your favorite python package manager.

It is easiest to get conan from pip by running

pip install conan==1.58

Install CMake from Here!

Next make a build directory and install with conan, and then build.

mkdir build
cd build
conan install --build missing ..
conan build ..

Conan Options

• cuda_arch is an option to specify the SM architecture you want to compile for, by default we compile for sm70 to sm90
• cuda_compiler is an option to specify the CUDA compiler for example nvcc

Code Organization

• include/lslab contains all of the lslab code
  • lslab.h contains basic macros
  • map.h contains a GPU interface for the map
  • hash.h contains GPU hash functions
  • device_allocator.h contains a device allocator
  • mutex.h contains a mutex implementation
  • set.h contains the set implementation
  • warp_mutex.h contains a cooperative warp mutex
  • detail contains extra implementation details
• test contains tests
• benchmark contains benchmarks

Clang Support

Clang seems to have an issue compiling the code that makes the tests fail. I am not going to fix this as of this moment unless necessary.