Global Trees

Global Trees and Global Chunks are frameworks for supporting global view programming of irregular data structures on distributed memory machines. This framework allows the user to create, update, and access distributed tree structures in a global fashion, using one-sided operations for access. This system allows for parallel computations on these data structures in an asynchronous manner. Communication operations are automatically mapped to bulk data movement, allowing for efficient fine-grained access of tree data.

Both Global Trees (GT) and Global Chunks (GCL) use ARMCI and MPI internally. We have tested interoperability with MPI, ARMCI, and Global Arrays applications and are investigating interoperability with additional parallel programming tools.

Contents 1 Scioto Programming Model 2 Documentation 2.1 Programming Interface 2.2 Publications 3 Project Members 4 Acknowledgements 5 Download

Scioto Programming Model

Scioto provides the user with a task pool programming model. Under this model the user expresses their computation as a set of tasks that can be executed in parallel. Tasks read inputs from their task descriptor as well as from data stored in the global address space. Tasks store their output in the global address space and are also permitted to add new tasks to the task collection.

Below is pseudocode that shows the structure of a simple Scioto program.

 // The user decides how to store task arguments within the task
 // body.  Generally, users define a struct to organize this data.
 typedef struct {
   double args;
 } mytask_t;

 // The user provides a function that implements the task.  This
 // function is run by Scioto when executing the task.
 void task_fcn(gtc_t gtc, task_t *my_task) {
   // ...
 }

 int main() {
   gtc_t    tc   = gtc_create(...);      // Create a task collection
   task_t   task = gtc_task_create(...); // Create a task descriptor
   mytask_t my_t = gtc_task_body(task);  // Get the task body
   
   my_t.args = ...;    // Fill in the task's arguments

   gtc_task_add(task); // Add the task to the task collection

   gtc_process(tc);    // Enter task-parallel region
}

Documentation

Programming Interface

Documentation on Scioto's user API can be found here: Scioto API

Publications

Scalable Work Stealing PDF PPT
James Dinan, Sriram Krishnamoorthy, D. Brian Larkins, Jarek Nieplocha, P. Sadayappan
Proc. 21st Intl. Conference on Supercomputing (SC). Portland, OR, Nov. 14-20, 2009.

Scioto: A Framework for Global-View Task Parallelism PDF PPT
James Dinan, Sriram Krishnamoorthy, D. Brian Larkins, Jarek Nieplocha, P. Sadayappan
Proc. of 37th Intl. Conference on Parallel Processing. Portland, OR, Sept. 8-12, 2008.

Project Members

• James Dinan (Scioto maintainer)
• D. Brian Larkins
• Prof. P. Sadayappan
• Sriram Krishnamoorthy
• Jarek Nieplocha

Acknowledgements

This research was supported in part by DOE grant #DE-FC02-06ER25755 and NSF grant #0403342.

Download

We are currently providing early development snapshots for download. Please see the included README file for installation instructions. For questions or comments, contact Brian Larkins (blarkins at coastal dot edu).

• Dev Release, 0.2 - July 2010 - [1]