Systems for Graph Neural Networks

Representation learning is a fundamental task in machine learning. It consists of learning the features of data items automatically, typically using a deep neural network (DNN), instead of selecting hand-engineered features that typically have worse performance. Graph data requires specific algorithms for representation learning such as DeepWalk, node2vec, and GraphSAGE. These algorithms first sample the input graph and then train a DNN based on the samples. It is common to use GPUs for training, but graph sampling on GPUs is challenging. Sampling is an embarrassingly parallel task since each sample can be generated independently. However, the irregularity of graphs makes it hard to use GPU resources effectively. Existing graph processing, mining, and representation learning systems do not effectively parallelize sampling and this negatively impacts the end-to-end performance of representation learning.

We developed NextDoor, the first system specifically designed to perform graph sampling on GPUs. NextDoor introduces a high-level API based on a novel paradigm for parallel graph sampling called transit-parallelism. We implement several graph sampling applications, and show that NextDoor runs them orders of magnitude faster than existing systems

Our current work is on split-parallelism, a novel hybrid parallelism strategies for mini-batch training. The traditional data parallel approach poses inherent limitations to scalable GNN training because it introduces redundant data loading and computation. Rather than proposing incremental patches to that paradigm, as done by previous work, this project proposes a fundamental shift towards a new paradigm for GNN training called split parallelism. The main idea is to split the computation of mini-batches across multiple GPUs, which cooperatively perform each iteration. Split parallelism scales the size of the cache when the number of GPUs in the system grows and maximizes cache access locality, which results in much better performance also in smaller models. We implemented split parallelism in a system called GSplit.