Project

General

Profile

Feature #1666

Support #1586: OpenMP integration support for external library written in and compiled by OpenMP on Charm++

Standalone OpenMP implementation based on Converse User-level threads

Added by Seonmyeong Bak about 1 month ago. Updated 19 days ago.

Status:
New
Priority:
Normal
Category:
-
Target version:
-
Start date:
09/06/2017
Due date:
% Done:

0%


Description

The current OpenMP implementation only works with Charm++ and AMPI.
However, this OpenMP implementation basically can work standalone because it only uses converse APIs which are independent of Charm++ and AMPI.

Several things should be done for this but it doesn't require huge changes in the current implementation.

1. Move the task queue into converse level
The current task queue is used only for OpenMP and there's no reason that it should be in ck-core. It only uses converse constructs.

2. Initialization of Converse library in OpenMP runtime.
LLVM OpenMP runtime has explicit initialization point so we only need to call 'ConverseInit' there and 'CsdExitScheduler' in the termination of the scheduler.

This work can bring many benefits other than the modularity of the OpenMP implementation.

1. We can accelerate all existing OpenMP applications with our work.
Because the OpenMP works on converse and can use all the feature of the converse, we can provide better OpenMP implementation
compared to the vanilla LLVM OpenMP and other OpenMP implementations including research OpenMP implementations such as ForestGOMP and BOLT.
In current implementation, we can accelerate only MPI + OpenMP applications other than Charm++ applications. However, MPI and Charm++ are only used for scientific applications, which have very limited pool of users.
If we want to see some applications we haven't touched, we should be able to run other parallel apps beyond MPI and Charm++.
OpenMP has bigger user pool and been used beyond academia and scientific applications such as machine learning and data analysis apps directly or indirectly through libraries.

2. OpenMP with other high-level languages and parallel frameworks.
There are many parallel frameworks in high-level languages which adopts fork-join model instead of message passing model.
We can accelerate these frameworks with our OpenMP in many ways. Even though we're working on python support of Charm++, it just ports Charm++ constructs into python and still highly complex to use for ordinary users.


Subtasks

Feature #1670: Migration of Taskqueue from Ck to Converse layerImplementedSeonmyeong Bak

History

#1 Updated by Seonmyeong Bak about 1 month ago

  • Assignee set to Seonmyeong Bak

#2 Updated by Seonmyeong Bak 20 days ago

  • % Done changed from 0 to 50

Basically, it works with a few lines of codes in converse and OpenMP runtime.

Also available in: Atom PDF