ROCm Discussions

[tetsuro]

I just started learning about ROCm capable GPU and inter-gpu communication.
I'm sorry, but could you help me resolve my question?

1. I cant find the doc what interconnected GPUs via Infinity Fabric Link can do.

Does these work as one GPU?
Or simply work as multiple GPUs?

2. I've seen the following two libraries for inter-node communication, which one is better?

MPI
rocALUTION

Or if neither one is good, what else is there?

[ValentinDerkovsky · Accepted solution]

With the advent of the Internet, distributed computing systems (DSS) began to develop. Currently, the most common architecture of DBC is a client-server. This architecture allows to transfer of the computational load server, and the client only displays the result. The resources of one server are finite, and in scalable projects, the problem of data exchange between processes, running on different processors or servers, arises in connection with the presence of a large number of processes. The authors cover the methods of designing client-server applications based on three-link architecture using different variants of interprocess communication. Interprocess communication methods in projects with client-server architecture and horizontal scaling are considered in this paper. The authors have made a comprehensive analysis of interprocess communication problems and proposed solution methods. The methods for organizing data exchange between server processes, client processes, and server processes are investigated in this paper. Authors have developed a methodology for performance testing of the system based on the main RTS paradigms and modern technologies. Based on the results the importance of interprocess communication design on all levels of three-link architecture was marked by the authors.