2019-07-25, 11:00–12:00, BoF: Room 353
This BoF will be a forum to discuss the state of the state around performant parallelism for distributed memory programming in Julia. Performance, parallelism, productivity and portability are four P's of distributed memory parallelism that over the last 30 years have proved hard to satisfy simultaneously in a general solution. The goal of this BoF is discussion and exploration of approaches for providing performant distributed memory parallelism in Julia in ways that are portable and that reflect the productivity vision of Julia. The format will consist of a series of presentations and a discussion/Q&A section. It will look both within Julia and across other languages at the last 30 years of efforts in this space. The motivation for the BoF is that meeting the four P's well remains an unsolved problem. For now projects that seek all of performance, parallelism at scale, portability and productivity typically have to make compromises in one or more of these areas. The hoped for outcome is some shared momentum and sharing of ideas for developing Julian approaches that lessen (or eliminate) the need to compromise in any of the four P's in the future.
This BoF is motivated in part by interesting new application development efforts that are combining high-performant scientific computing style parallelism with large scale machine learning, optimization and statistical methods. The BoF will focus on current and future directions for language and library level abstractions for large distributed parallel applications in Julia. Some projects such as Celeste and CLiMA are employing Julia for work that scales to tens of thousands or more parallel cores. The BoF will be a forum for starting conversations and brainstorming how Julia tools for distributed parallelism might evolve in coming years to support both portable, high-performance parallelism and productive interactivity in a relatively unified way. The BoF will include short presentations from current participants in large scale scientific computing parallel efforts using Julia, it will also include short presentations from relevant tool developers as well as an open discussion time for question and answer. We anticipate covering a few interconnected themes and goals.
Blending in new ideas. In the last 30 years, since the emergence of the message passing interface (MPI) parallel library standard for scientific computing, many projects have explored alternate parallelism paradigms. These alternates, which include UPC, X10, CoArray Fortran, Chapel, Hadoop, Spark, OCCA and many others, have introduced new ideas, but in scientific computing none have gained the broad traction of MPI in practice. One topic of interest for this BoF is where can the Julia language, library and meta-programming ecosystem usefully blend in some of the parallel programming ideas that have emerged in academic projects in the last 30 years.
Enabling interactivity front and center in large scale, high-performance parallelism. Are there ways Julia might help bring some flavor of productive interactivity to the sorts of applications that have typically leveraged MPI for portable performance. Some debugging tools, from other languages, such as ddt and totalview may provide models for how to interact explicitly with multiple concurrent streams of computation in a large parallel application under something of a REPL feel. Julia Distributed arrays, Channels and Cluster manager abstractions provide more implicit ways for interacting with parallel computation streams. Using existing components it is already possible to create elaborate applications that mix high-performance parallel simulation and distributed machine learning, for example, in a single workflow. This sort of workflow is emerging in many fields from plasma science to economics. However, current programming approaches can result in a somewhat complex mix of abstractions that are not always amenable to flexible and agile exploration in an interactive environment of the sort that is Julia's hallmark.
Catalyzing activities to discover the right APIs. The BoF will aim to catalyze conversations and energize projects looking at the next generation of Julia ecosystem tools to support high-performance, parallel scientific computing. In this realm it is likely that many of the key abstractions/APIs are likely to be discovered somewhat organically as much as designed. The BoF seeks to be inclusive to all Julia projects that may have an interest in these areas. The organizers welcome anyone interested in presenting a slide at the BoF to contact us. We plan to gather the BoF material into a post meeting document that will be openly available.