Use io_cache for Infer action inputs and outputs
This issue comes after #5 (closed) once the Worker network handler is implemented.
Infer action input and Infer response output are implemented as message payloads.
Ideally this memory is not allocated using malloc
or cudaMallocHost
, but instead re-uses memory that has already been allocated with the Memory Manager's IOCache
defined in memory.h. The clockwork worker's IOCache can be accessed at worker->runtime->manager->io_cache