Distributed systems incorporate GPUs because they provide massive parallelism in an energy-efficient manner. Unfortunately, existing programming models make it difficult to route a GPU-initiated network message. The traditional coprocessor model forces programmers to manually route messages through the host CPU. Other models allow GPU-initiated communication, but are inefficient for small messages.To enable fine-grain PGAS-style communication between threads executing on different GPUs, we introduce Gravel. GPU-initiated messages are offloaded through a GPU-efficient concurrent queue to an aggregator (implemented with CPU threads), which combines messages targeting to the same destination. Gravel leverages diverged work-group-level semantics to amortize synchronization across the GPU's data-parallel lanes.Using Gravel, we can distribute six applications, each with frequent small messages, across a cluster of eight GPU-accelerated nodes. Compared to one node, these applications run 5.3x faster, on average. Furthermore, we show Gravel is more programmable and usually performs better than prior GPU networking models.
