We have three principal near-term priorities for Gunrock development (as of June 2019):
Gunrock's 1.0 release focused on single-GPU performance. We expect to target single-node multi-GPU support in a near-term future release.
We believe dynamic (mutable) graphs are a practically interesting area of graph analytics that has not been a focus of the GPU graph analytics research community.
We will ensure that Gunrock will integrate with NVIDIA's RAPIDS suite of software libraries to accelerate data science workloads on GPUs.
We also continue to pursue projects in scalability (within a node, across nodes, and across the CPU-GPU boundary); in programmability, including working with the MIT GraphIt team; in improving Gunrock's internals, both their performance and their modularity; in expanding the number and coverage of applications written in Gunrock; and in better interoperability with other tools, languages, and libraries.