This work is part of a research project that investigates index structure accesses over fast interconnects. We thereby assume that index structures are stored out-of-core and accessible through interconnect technologies like NVLink 2.0.

The project itself consists of a series of carefully designed microbenchmarks. Each of the benchmarks aims to highlight or respectively solve the challenges imposed by the data transfer bottleneck.

Requirements

• CUDA-capable device with device architecture 7.0 or higher
• NVIDIA CUDA toolkit/compiler version $\ge$ v11.2
• C++14 or higher

Building

mkdir -p build
cd build
cmake -DCMAKE_CUDA_FLAGS="-gencode=arch=compute_70,code=sm_70" ..
make

The gencode argument should be adapted accordingly to reflect the target GPU's capabilities.

Benchmarking

After the build process succeeded all binaries required by the benchmark scripts can be found in the previously created build directory. The benchmark scripts themselves are located in scripts.

In order to reproduce our results, it is sufficient to execute the corresponding benchmark script. If, for example, we want to recreate throughput results, we would run the following from within the build directory:

../bench_join_throughput.sh

This will generate a .yml file in the current directory. It contains information like the execution time, and parameters such as the used memory allocators.

Note: Certain parameters may require adjustment in order to reflect the target machine's environment.

Research

We plan to publish our findings from this project.