PMAW: The Programming Models and Algorithms Workshop

Workshop Overview

Programming models are widespread, and have gained vast popularity due to a variety of optimization mechanisms and abstractions, their transparency and other runtime features. Examples of these features are automatic accelerator offloading, directive-based compiler optimizations and parallelization, integration into high-level programming languages, locality enhancements, work stealing and pushing, overlapping communication and computation, providing a wide range of dynamic scheduling techniques, scaling and enabling the execution of large scale applications for distributed computing, leveraging and exposing interfaces that enable a fine grained control of underlying computing and network hardware, among many other features. Despite this long list of major advancements, there still exists a wide gap between the algorithmic design, specification and implementation, and these programming models.

To exacerbate the problem at hand, it is widely known that to achieve the expected levels of performance and power cap of Exascale and future computing eras, tremendous efforts will have to be invested and will require the convergence of several traditionally disjoint fields. Programming models excel at raising the level of abstraction from a concrete target platform, architecture or even the underlying network. In other words, their primary role is to hide the complexities of some lower level of the hardware and software stack, thereby greatly simplifying and speeding the development process of scientific software.

Complicating even more the matter, new technologies such as upcoming programming models and parallel runtimes are not fully exploited. To a large extent, this is caused by the unfamiliarity of domain experts and computational scientists with new runtimes and software developments, or conversely, because of the lack of knowledge of the runtime or compiler expert with the fundamental properties of a computational algorithm. Moreover, simple mechanical porting of traditional and classic algorithms, many of which were devised 40 or more years ago, considered much more generic and abstract models such as the P-RAM model, the Bulk-Synchronous Parallel (BSP) model, LogP, or even the traditional von Neumann model. More precisely, this process has been target agnostic, and did not envision any specific hardware or runtime. This approach often results in suboptimal performance, loses important information in translation or incorrectly exploits the algorithmic intrinsics.