TinyMPI: A virtualized MPI Implementation for Arm and X86
Motivation
Traditional MPI implementations run at most one MPI rank per CPU core.
TinyMPI runs more than one MPI rank per CPU core, i.e., it oversubscribes the
CPU, with the goal of achieving automatic computation-communication overlap:
when one MPI rank blocks, TinyMPI switches to another and continues using the
CPU.
We have used TinyMPI as a tool in the research effort of answering the
question, "How many ranks exactly to start on each CPU core?", the results of
which - in the form of a model - are presented here.
Performance Model
The number of MPI ranks that a virtualized MPI implementation launches per CPU
core is referred to as virtualization ratio. We will also refer to
this parameter as V. Intuitively, V plays an important role in the overall
performance: if V=1, then no virtualization benefits can be seen at all; if
V is too large, then the overhead of task switching and the increased volume
of intra-node communication will outweigh all benefits. The project report contains
a full description of our performance model.
Implementation
References
