Architecture

DeviceDriver

The manager of all the GPU resources, responsible for allocating and destroying resources (e.g.: pointers, streams, events). It also provides primitives for interacting with these GPU resources. However, it has no domain knowledge about the MASS APIs.

Each GPU participating in the simulation is assigned with one DeviceDriver.

DeviceDispatcher

The component containing all the kernel implementations and host-side logic for invoking these kernels. Each DeviceDispatcher manages one instance of DeviceDriver and is responsible for one GPU.

It is not responsible for the actual data transfer between GPUs, so it has no knowledge about the other GPUs and the other DeviceDispatcher instances. However, it does have information about the existance of its neighbors at certain directions.

RootDispatcher

One instance for each host and maintains all the DeviceDispatcher instances for the GPUs on that host. It is responsible for dispatching the MASS CUDA APIs requests to all the DeviceDispatcher instances and handling all the actual data transfers between GPUs.

MASS APIs

The component containing all the functions, types and interfaces for interacting with the library.