This paper presents an implementation of the new endpoint proposal in the message passing interface (MPI) standard. The objective of the proposal is to allow threads to efficiently access network resources and communicate. The paper proposes EP-lib, a library-based implementation that can be used by any MPI library to support the endpoint proposal. The main concept of the library uses additional processes (proxies) that intercept endpoint-based communication and efficiently access the network and do the transfer. Using an Ohio State University (OSU) micro-benchmark as well as a lattice quantum chromodynamics (QCD) application kernel and a fast Fourier transform (FFT) communication benchmark, the authors evaluate the impact of the new design on the hybrid MPI+OpenMP programming model’s performance. Further, the evaluation shows that the overhead of EP-lib is a negligible constant value. This is mainly due to an efficient implementation that maps the user space data to the proxy space.

The paper is well written and flows very well. The authors discuss their designs with regard to the different MPI operations and their semantics. The paper’s contributions and comparisons to related work are highlighted.

I have a few concerns with the paper that are not exhaustively discussed:

(1) As jobs are automatically changed to surrogate jobs (with more processes), how are the processes computed?

(2) Can the number of processes change dynamically during execution if the number of OpenMP threads varies between the different sections?

(3) What exactly is the mapping between the number of threads and proxies? If the proxies are multithreaded, how do they efficiently access the network without sharing structures?