Performance portability, in the sense that a single source can run with good performance across a wide vari-ation of parallel hardware platforms, is strongly desired by industry and actively being researched. How-ever, evidence is mounting that performance portability cannot be realized at just the toolchain level, or just at the runtime level or just at the hardware abstraction level. This is a
... [Show full abstract] position paper, making a suggestion for how the groups involved can more efficiently solve the performance portability problem together. We don't propose a solution, at all, but rather a support system for the players to self organize and collectively find one. The support system is based on a new extendable virtualization mechanism called VMS (Virtualized Master-Slave), that fulfills the needs of an organizing principle, and provides focus that may increase research efficiency. The difficult work will be the on-going research efforts on parallel language design, compilers, source-to-source transform tools, binary optimization, run-time schedulers, and hardware support for parallelism. Although it doesn't in itself solve the problem, such an organizing principle may be a valuable step towards a solution – the problem may be too complex and require cooperation of too many real-world entities for a single-entity solution. We briefly review VMS, and illustrate how it could be used to give rise to an eco-system in which perfor-mance portability is collectively realized. To support the suggestion, we give measurements of the time to implement three parallelism-construct libraries, and performance numbers for them, along with measure-ments of the basic overhead of VMS.
