REMOTE GPU OFFLOADING IN CMSSW

The intention of this report is to give an overview both about the outcome of our project and the approach we have followed in order to reach it. The main achievement has been the implementation in the CMS Software of a way to perform Remote Direct Memory Access (RDMA) directly on the GPU of a remote server. The underlying motivation is the desire to offload work to the GPU of a remote machine in order to free the CPU resources of the local host, which can therefore be used for other purposes while the computation is taking place remotely. The goal is not only to use RDMA to make this communication possible, but also to do it with a good compromise between the overall latency and the CPU load on the client side.