About plugins🔗

Plugins are an essential part of vAccel. They contain the implementations of the operations defined in the API. By keeping the implementation logic separate from the operation definition, vAccel enables the use of different acceleration frameworks without requiring the modification or recompilation of the application binaries. This means, that a compiled vAccel application can use any acceleration implementation available as a plugin by simply specifying the plugin at runtime.

A vAccel plugin has the form of a shared object (library) that can be loaded with dlopen(). The object can be compiled from any language as long as the contained symbols/functions are callable from C.

You can better understand the use of the plugin system by trying to run an example with different plugins.

The image classification example is based on the relevant User API operation. The operation is supported by the NoOp plugin, that we used in Running the examples. It is also supported by the Torch and the TVM plugins. This means that the same classify binary can be used to run with the NoOp, the Torch and the TVM vAccel image classification implementations simply by specifying the relevant plugin (provided all dependencies are installed).