XPRESS: eXascale PRogramming Environment and System Software Sandia National Laboratories PI: Ron Brightwell Thomas Sterling, Project Chief Scientist, Indiana University Barbara Chapman, University of Houston Hartmut Kaiser, Louisiana State University Alice Koniges, Lawrence Berkley National Laboratory Andrew Lumsdaine, Indiana University Allen Maloney, Oregon University Alan Porterfield, University of North Carolina Chris Baker, Oak Ridge National Laboratory

1  

Scope  of  the  Project   New Model Applications

Legacy Applications

OpenMP

Domain Specific Language Domain Specific Active Library

MPI

Metaprogramming Framework

2. LXK lightweight OS for O(N) scalability, performance portability, no OS noise and minimum abstraction layer to runtime software and billion-way hardware

Compiler XPI

Operating System Instances

{ {

Hardware Architecture

AGAS

LCO

name space processor

dataflow, futures synchronization

Parcels Lightweight message driven Threads context manager

computation

PRIME MEDIUM Interace / Control

Operating System

...

Distributed Framework Address space control

Task recognition Memory bank control

OS thread

Network drivers

Instrumentation

Runtime System Instances

...

1. Guided by ParalleX execution model & SLOW performance model for codesign of dynamic adaptive resource management and task scheduling system software & programming model with introspection

+106 nodes ⇥ 103 cores / node + integration network

3. HPX runtime software for efficiency and scalability of science application by adapting workload to resource availability and discovery of datadirected parallelism 4. XPI & RIOS interface protocols for cross-platform portability

Scope  of  the  Project   v Establish logical and organizational abstractions for foundations of general-purpose exascale computing architectures and programming models v Enable scalability to exascale by dynamic parallelism discovery and low overhead for fine grain threads v Achieve best time to solution and low energy by up to 10X efficiency for real apps through low overhead and adaptive latency and contention management v Full-scale deployment of prototype XPRESS system software for general-purpose DOE applications v Legacy codes and methods migration path v Partnering with other DOD, DOE, NSF programs

Milestones  &  Deliverables  

•  Milestones  &  deliverables   1.  2.  3.  4.  5.  6. 

XPI  intermediate  form  and  low  level  programming  interface   RIOS  interface  protocol  between  OS  and  run?me  system   LXK  lightweight  kernel  scalable  opera?ng  system   HPX-­‐4  run?me  system  based  on  ParalleX  execu?on  model   MPI  transla?on  to  XPI  for  legacy  code  migra?on   ParalleX,  introspec?on  framework  and  policies,  translators   Milestones (Level 1) •  XPI & RIOS Specifications •  Merger of Kitten & HPX-3 •  Introspective framework •  MPI to XPI Port •  HPX-4 Developed with LXK •  Deploy, Test, Evaluate with apps

Scheduled FEB 2013 AUG 2013 FEB 2014 AUG 2014 FEB 2015 Aug 2015

Dependencies   •  Needs  from  other  projects?     –  –  –  –  –  – 

Which  projects  do  you  think  can  provide  what  you  need   3  co-­‐design  centers  for  applica?on  drivers   Quinlan  project  for  DSL   OCR  for  high  level  programming  interface   Intel  project  for  possible  target  architecture   MIT  for  control  policies  

•  What  we  offer  to  other  projects?   –  –  –  –  – 

Lightweight  kernel  opera?ng  system  alterna?ve  to  Linux   Run?me  system  that  is  dynamically  adap?ve  and  scalable   Low  level  intermediate  form  and  programming  interface   Protocol  specifica?on  for  interface  between  OS  and  Run?me   ParalleX  execu?on  model  abstrac?on  for  seman?c  founda?on  

Gaps   •  Gaps  in  your  project     –  Interface  to  co-­‐design  centers;  Post-­‐doc  model  may  not  work   –  Where  is  the  parallelism  going  to  come  from  to  drive  billion-­‐wary   concurrency   –  Performance  models  for  exascale  system   –  Well  founded  technology  roadmap  to  base  2020  extrapola?ons  

•  An?cipated  needs  in  the  future   –  –  –  – 

Testbed  plaWorms  for  proof-­‐of-­‐concept  test  and  cer?fica?on   Base-­‐level  benchmarks  and  metrics   XPI  versions  of  key  applica?on  kernels  with  data  sets   I/O  models  including  mass  storage  for  big  data