Open Community Runtime

Introduction

Goal

The goal of the Open Community Runtime (OCR) project is to propose and evaluate an API and framework aimed at:

• Expressing large amounts of parallelism in a task-based model
• Explicitly representing data dependences
• Running efficiently on tomorrow's exascale machines

Audience

In its current state, the project aims to release a prototype implementation of a reference API by September 2015. As such, the project is mostly geared towards early adopters who would like to provide feedback on our programming model and API, runtime developers who are interested in implementing various algorithms and heuristics within our runtime framework and higher-level language/models implementors who are interested in determining if their model can map to OCR.

The initial release of OCR will be a development platform aimed at better understanding the challenges of executing task-based programming models on large scale systems.

Value of the OCR project

The OCR project is creating an application building framework that explores new methods of high-core-count programming with an initial focus on HPC applications. The project aims to explore, among other things:

• Expressability: how can application programmers express their applications in a hardware-agnostic manner; i.e.: how can they express the semantic of an application as opposed to the mapping of it to a particular kind of hardware
• Scheduling and data placement: what heuristics allow locality-aware scheduling and data-placement on an exascale system
• Introspection: how can a runtime system improve its execution characteristics by monitoring itself
• Resiliency: how can a runtime system deal with failures in an exascale system

The OCR project aims to propose a low-level API to address the challenges of exascale programming and is meant to be targeted by higher-level abstractions.

Timeline

The OCR software is available under the BSD open source license.

• Initial unveiling at SC 2012
• v0.8 was introduced in SC 2013 and was a significant rewrite of most of the internal code to increase modularity and allow more community participation
• v0.9 was released in mid October 2014.
• An initial draft of the specification will be released at SC 2014

Links

• The source code is available on GitHub
• Mailing lists are available here

Install Instructions for Distributed-OCR

The 'sc14' github branch contains the most recent OCR implementation publicly available. Things will be streamlined along the way.

Checking out the code

git clone https://github.com/01org/ocr.git
cd ocr
git branch sc14 --track origin/sc14
git checkout sc14

Environment setup

From the checked-out ocr folder setup the following environment variable.

export OCR_INSTALL_ROOT=$PWD/install
export OCR_SRC=$PWD
export OCR_BUILD_ROOT=$PWD/build
export APPS_ROOT=$PWD/ocr-apps

Compiling and Running Applications

The ocr-apps folder contains various applications. Those applications relies on pre-defined makefiles to ease compiling and linking against the OCR interface.

The general pattern to get an OCR application to run is as follow:

• Select a target platform
• Invoke the corresponding Makefile to build the application
• Select an OCR configuration file
• Invoke the corresponding Makefile to run the application

For instance, the npbCG application defines Makefiles for the following targets:

• x86-pthread-x86: for shared-memory OCR
• x86-pthread-mpi: for distributed-memory OCR relying on MPI for communications.

Hence to build npbCG for distributed, one executes the following command:

make -f Makefile.x86-pthread-mpi prerun install

Note that the Makefile checks if the runtime needs to be built too.

Once the application is built for distributed a configuration file must be provided. OCR relies on text-based configuration files to setup runtime instances. Some sample configuration files are available under machine-configs/

export OCR_CONFIG=/path/to/configfile

Additionally, one can point to file containing hostname to use for the the distributed run

export OCR_NODEFILE=/path/to/nodefile

Running the application can be done through the Makefile system as well:

make -f Makefile.x86-pthread-mpi run WORKLOAD_ARGS"-t B -b 1000"