HPX-5: Difference between revisions

Revision as of 18:25, November 20, 2015

Announcement

HPX-5 version 2.0 Runtime System Release Announced by Indiana University!

The Center for Research in Extreme Scale Technologies (CREST) at Indiana University is pleased to announce the release of version 2.0 of the HPX-5 runtime system for petascale/exascale computing. Building on CREST’s commitment to developing new approaches for achieving the highest levels of performance on current and next-generation supercomputing platforms, HPX-5 is provided to support the international high-performance computing community in addressing significant challenges involved in achieving exascale computing.

HPX-5 is a reduction to practice of the revolutionary ParalleX execution model, which establishes roles and responsibilities between layers in an exascale system. It is implemented in portable C99 and is organized around a multi-threaded, cooperative thread scheduler, a global address space, an active-message parcel transport, and a group of globally addressable local synchronization object classes. Internally, the infrastructure is built on scalable concurrent data structures to minimize shared-memory synchronization overhead. The global address space and parcel transport are based on the innovative Photon network transport library, which supports low-level access to network hardware and provides RDMA with remote completion events for low overhead signaling. An alternative ISend/IRecv network layer is included for portability, along with a reference MPI implementation. HPX-5 is compatible with Linux running on Intel x86 and Xeon Phi processors and various ARM core platforms (including both ARMv7 and ARMv8/Aarch64). A pre-release version of HPX-5 v2.0 is available for Mac OS X 10.10+.

“HPX-5 is a useful environment for exploring dynamic adaptive execution for high scalability computation as well as critical support for truly dynamic end-user science and engineering problems” said Thomas Sterling, professor at Indiana University and creator of the ParalleX execution model.

Introduction

HPX-5 v2.0 (High Performance ParalleX) is the latest version of the HPX-5 runtime system that provides a unified programming model for parallel and distributed applications, allowing programs to run unmodified on systems from a single SMP to large clusters and supercomputers with thousands of nodes. HPX-5 is a realization of ParalleX, an abstract cross-cutting exascale execution model, which establishes roles and responsibilities between system layers.

The HPX-5 interface and library implementation is guided by the ParalleX execution model and is freely available, open source, portable, and performance-oriented. HPX-5 is a general-purpose runtime system for applications, targeted at conventional, widely available architectures. It provides a unified programming model for parallel and distributed applications. As a dynamic adaptive runtime system, it is event-driven and embodies the principles of multi-threaded computing while also providing a global name and address space and advanced synchronization constructs. For communication between nodes, it uses the Photon network layer which has been tuned for optimal single sided communications as an alternative to the Message Passing Interface (MPI).

HPX-5 Architecture

The ParalleX execution model itself is experimental and continues to evolve as driven by quantitative insights gathered from the Starvation-Latency-Overhead-Waiting for contention (SLOW) performance model. It aims to address the aggravating effects of asynchrony for extreme-scale machines and to improve the efficiency and scalability of scaling constrained applications. ParalleX utilizes lightweight concurrent threads managed using synchronization primitives such as dataflow and futures in order to alter the application flow structure from being message-passing to becoming message-driven. It also includes an advanced global address space model instead of relying on more conventional distributed memory structures.

Audience

HPX-5 is used for a broad range of scientific applications, helping scientists and developers write code that shows better performance on irregular applications and at scale when compared to more conventional programming models such as MPI. For the application developer, it provides dynamic adaptive resource management and task scheduling to reach otherwise unachievable efficiencies in time and energy and scalability. HPX-5 supports such applications with implementation of features like Active Global Address Space (AGAS), ParalleX Processes, Complexes (ParalleX Threads and Thread Management), Parcel Transport and Parcel Management, Local Control Objects (LCOs) and Localities. Fine-grained computation is expressed using actions. Computation is logically grouped into processes to provide quiescence and termination detection. LCOs are synchronization objects that manage local and distributed control flow and have a global address. The heart of HPX-5 is a lightweight thread scheduler that directly schedules lightweight actions by multiplexing them on a set of heavyweight scheduler threads.

Features in HPX-5

Fine grained execution through blockable lightweight threads and unified access to a global address space.
High-performance PGAS implementation which supports low-level one-sided operations and two-sided active messages with continuations, and an experimental AGAS option that allows the binding of global to physical addresses to vary dynamically.
Makes concurrency manageable with globally allocated lightweight control objects (LCOs) based synchronization (futures, gates, reductions) allowing thread execution or parcel instantiation to wait for events without execution resource consumption.
Higher level abstractions including asynchronous remote-procedure-call options, data parallel loop constructs, and system abstractions like timers.
Implementation of ParalleX processes providing programmers with termination detection and per-process collectives.
Photon networking library synthesizing RDMA-with-remote-completion directly on top of uGNI, IB verbs, or libfabric. For portability and legacy support, HPX-5 emulates RDMA-with-remote-completion using MPI point-to-point messaging.
Leverages distributed GPU and co-processors (Intel Xeon Phi) through experimental OpenCL support.
PAPI support for profiling. APEX policy engine (Autonomic Performance Environment for eXascale) support for runtime adaption.
Migration of legacy applications through easy interoperability and co-existence with traditional runtimes like MPI. HPX-5 2.0 is also released along with several applications: LULESH, Wavelet AMR, HPCG and the ParalleX Graph Library.

Timeline

The HPX-5 source code is distributed with a liberal open-source license.

v1.0.0 released on 3rd May 2015.
v2.0.0 released on 17th November 2015.
v3.0.0 is scheduled to be released in April/May 2016.

HPX-5 is developed with an agile process that includes continuous integration, regular point releases, and frequent regression tests for correctness and performance. Users can submit issue reports to the development team through the HPX-5 web site(https://gitlab.crest.iu.edu/extreme/hpx/issues). Future major releases of HPX-5 will be delivered semi-annually although bug fixes will be made available between major releases.

The HPX-5 source code is released under the BSD open-source license and is distributed with a complete set of tests along with selected sample applications. HPX-5 is funded and supported by the DoD, DoE, and NSF, and is used actively in projects such as PSAAP, XPRESS, XSEDE. Further information and downloads for HPX-5 can be found at http://hpx.crest.iu.edu.

Quick Start Instructions

If you plan to use HPX–5, we suggest to start with the latest released version (currently HPX–5 v2.0.0) which can be downloaded from https://hpx.crest.iu.edu/download. Follow the installation directions located under hpx/INSTALL

Links

Webpage: http://hpx.crest.iu.edu
HPX-5 Documentation (http://hpx.crest.iu.edu/documentation). User guide and developers guide are online, along with HPX-5 developers policy
HPX-5 tutorial is available at (http://hpx.crest.iu.edu/tutorials).
HPX-5 Source code (http://hpx.crest.iu.edu/git_repository) The HPX-5 source code is available as a git repository to be cloned.
Applications (http://hpx.crest.iu.edu/applications).
Frequently asked questions is available at (http://hpx.crest.iu.edu/faqs_and_tutorials).
Mailing lists are available here: http://hpx.crest.iu.edu/contact.

@@ Line 2: / Line 2: @@
 == Announcement ==
-HPX-5 v1.0.0 released! Links
+'''''HPX-5 version 2.0 Runtime System Release Announced by Indiana University!'''''
+The Center for Research in Extreme Scale Technologies (CREST) at Indiana University is pleased to announce  the release of version 2.0 of the HPX-5 runtime system for petascale/exascale computing. Building on CREST’s commitment to developing new approaches for achieving the highest levels of performance on current and next-generation supercomputing platforms, HPX-5 is provided to support the international high-performance computing community in addressing significant challenges involved in achieving exascale computing.
+HPX-5 is a reduction to practice of the revolutionary ParalleX execution model, which establishes roles and responsibilities between layers in an exascale system. It is implemented in portable C99 and is organized around a multi-threaded, cooperative thread scheduler, a global address space, an active-message parcel transport, and a group of globally addressable local synchronization object classes. Internally, the infrastructure is built on scalable concurrent data structures to minimize shared-memory synchronization overhead. The global address space and parcel transport are based on the innovative Photon network transport library, which supports low-level access to network hardware and provides RDMA with remote completion events for low overhead signaling. An alternative ISend/IRecv network layer is included for portability, along with a reference MPI implementation. HPX-5 is compatible with Linux running on Intel x86 and Xeon Phi processors and various ARM core platforms (including both ARMv7 and ARMv8/Aarch64). A pre-release version of HPX-5 v2.0 is available for Mac OS X 10.10+.
+“HPX-5 is a useful environment for exploring dynamic adaptive execution for high scalability computation as well as critical support for truly dynamic end-user science and engineering problems” said Thomas Sterling, professor at Indiana University and creator of the ParalleX execution model.
 == Introduction ==
-HPX-5 v1.0.0 (High Performance ParalleX) is the latest version of the HPX-5 runtime system that provides a unified programming model for parallel and distributed applications, allowing programs to run unmodified on systems from a single SMP to large clusters and supercomputers with thousands of nodes. HPX-5 is a realization of ParalleX, an abstract cross-cutting exascale execution model, which establishes roles and responsibilities between system layers.
+HPX-5 v2.0 (High Performance ParalleX) is the latest version of the HPX-5 runtime system that provides a unified programming model for parallel and distributed applications, allowing programs to run unmodified on systems from a single SMP to large clusters and supercomputers with thousands of nodes. HPX-5 is a realization of ParalleX, an abstract cross-cutting exascale execution model, which establishes roles and responsibilities between system layers.
+The HPX-5 interface and library implementation is guided by the ParalleX execution model and is freely available, open source, portable, and performance-oriented. HPX-5 is a general-purpose runtime system for applications, targeted at conventional, widely available architectures. It provides a unified programming model for parallel and distributed applications. As a dynamic adaptive runtime system, it is event-driven and embodies the principles of multi-threaded computing while also providing a global name and address space and advanced synchronization constructs. For communication between nodes, it uses the Photon network layer which has been tuned for optimal single sided communications as an alternative to the Message Passing Interface (MPI).
-The HPX-5 interface and library implementation is guided by the ParalleX execution model and is freely available, open source, portable, and performance-oriented. HPX-5 is a general-purpose runtime system for applications, targeted at conventional, widely available architectures. It provides a unified programming model for parallel and distributed applications. HPX-5 includes support for Linux on x86_64 and ARM architectures along with experimental support for Mac OS X. As a dynamic adaptive runtime system, it is event-driven and embodies the principles of multi-threaded computing while also providing a global name and address space and advanced synchronization constructs. For communication between nodes, it uses the Photon network layer which has been tuned for optimal single sided communications as an alternative to the Message Passing Interface (MPI).
+[[ File:Architecture.png|400x400px|frame|HPX-5 Architecture ]]
 The ParalleX execution model itself is experimental and continues to evolve as driven by quantitative insights gathered from the Starvation-Latency-Overhead-Waiting for contention (SLOW) performance model. It aims to address the aggravating effects of asynchrony for extreme-scale machines and to improve the efficiency and scalability of scaling constrained applications. ParalleX utilizes lightweight concurrent threads managed using synchronization primitives such as dataflow and futures in order to alter the application flow structure from being message-passing to becoming message-driven. It also includes an advanced global address space model instead of relying on more conventional distributed memory structures.
@@ Line 18: / Line 26: @@
 == Features in HPX-5 ==
-The HPX-5 C library implementation, libhpx, provides a high-performance, portable implementation of the API that runs on both SMP and distributed systems. It is designed around a cooperative lightweight thread scheduler and unified access to a global address space. Event driven programs invoke remote actions on global addresses using HPX-5’s parcel active-message plus continuation abstraction, or read or write global data directly through one-sided network operations. Globally addressable lightweight control objects (LCOs) provide control and data synchronization, allowing thread execution or parcel instantiation to wait for events without execution resource consumption. Finally, HPX-5’s implementation of ParalleX processes provides programmers the powerful abstraction of termination groups for parcel and thread execution. In addition to this core-programming model HPX-5 provides a number of higher level abstractions including asynchronous remote-procedure-call options, data parallel loop constructs, global memory allocation, and system abstractions like timers.
+*  Fine grained execution through blockable lightweight threads and unified access to a global address space.
+*  High-performance PGAS implementation which supports low-level one-sided operations and two-sided active messages with continuations, and an experimental AGAS option that allows the binding of global to physical addresses to vary dynamically.
-'''''Global Address Space:''''' The global address space is modeled as a linear byte addressable virtual space, where the binding between global addresses and local virtual addresses may change at runtime. Global addresses serve as the targets of active message parcels, as well as supporting put/get operations directly. HPX-5 ships with two implementations of the global address space. A traditional, high-performance partitioned global address space (PGAS) allows low-latency, one-sided put/get operations in addition to parcel access but does not support remapping, while an experimental active global address space emulates put/get with parcels while allowing the binding of global to physical addresses to vary dynamically.
+*  Makes concurrency manageable with globally allocated lightweight control objects (LCOs) based synchronization (futures, gates, reductions) allowing thread execution or parcel instantiation to wait for events without execution resource consumption.
+*  Higher level abstractions including asynchronous remote-procedure-call options, data parallel loop constructs, and system abstractions like timers.
-Memory management within the global address space leverages features of the high-performance jemalloc allocator, to provide scalable local allocation of globally addressable data, as well as cyclic distributed arrays.
+*  Implementation of ParalleX processes providing programmers with termination detection and per-process collectives.
-Lightweight Thread Scheduling: The HPX-5 lightweight-threading system enables massively threaded computation by minimizing the cost of thread creation and context switching. HPX-5 threads have unrestricted behavior and can wait for asynchronous events using lightweight control objects. Local load balancing is performed via workstealing, while distributed load balancing is managed by remapping in the active global address space. Threading overheads rival existing modern packages with compatible semantics.
+*  Photon networking library synthesizing RDMA-with-remote-completion directly on top of uGNI, IB verbs, or libfabric. For portability and legacy support, HPX-5 emulates RDMA-with-remote-completion using MPI point-to-point messaging.
+*  Leverages distributed GPU and co-processors (Intel Xeon Phi) through experimental OpenCL support.
-'''''Parcels:''''' HPX-5 parcels encode event-driven execution by extending a traditional active message abstraction with a continuation, enabling complex chains of computation in addition to traditional remote procedure calls. Parcels are sent to addresses within the global address space and embody two-sided network programming. Parcel instantiation is isomorphic with lightweight thread creation, in fact the parcel structure and thread descriptor are the same structure.
+*  PAPI support for profiling. APEX policy engine (Autonomic Performance Environment for eXascale) support for runtime adaption.
+*  Migration of legacy applications through easy interoperability and co-existence with traditional runtimes like MPI. HPX-5 2.0 is also released along with several applications: LULESH, Wavelet AMR, HPCG and the ParalleX Graph Library.
-'''''Local Control Objects (LCOs):''''' LCOs represent control state in memory. Both threads and parcels interact with LCOs, threads may block until an LCO is ready and/or its data is available, while parcel send operations may wait until an LCO is ready. HPX-5 is distributed with a number of built-in LCO types, including futures, and gates, generation counters, parameterizable commutative-associative reductions, semaphores, etc., and supports user-defined LCOs as well.
-Processes: HPX-5 groups parcels and threads into processes, which may optionally have termination detection enabled. Termination detection triggers an event when the processed has reached quiescence. This powerful mechanism can be useful in algorithms that do not contain a natural "join" point.
-'''''Networking:''''' In addition to active message parcel operations, HPX-5 provides direct access to the global address space through asynchronous put/get operations. Generic support for both parcel transport and put/get operations is designed around a novel networking abstraction, put-with-remote-notification. HPX-5 provides two implementations of this abstraction. The PWC network uses Photon’s put-with-remote-completion support directly and implements a parcel emulation layer for parcel transport. The ISIR network provides parcel transport on top of the traditional MPI Isend/Irecv point-to-point messaging, and emulates put-with-remote-notification with parcels.
-'''''Convenience Features:''''' Development of HPX-5 was guided by codesign with computational scientists. This resulted in a number of convenience features that encapsulate lower level behavior in more convenient and familiar form. These include a suite of remote-procedure style calls, data parallel loops for local and distributed data, and numerous smaller features.
 == Timeline ==
@@ Line 39: / Line 41: @@
 *  v1.0.0 released on 3rd May 2015.
-*  v2.0.0 is scheduled to be released at SC 2015.
+*  v2.0.0 released on 17th November 2015.
+*  v3.0.0 is scheduled to be released in April/May 2016.
+HPX-5 is developed with an agile process that includes continuous integration, regular point releases, and frequent regression tests for correctness and performance.  Users can submit issue reports to the development team through the HPX-5 web site(https://gitlab.crest.iu.edu/extreme/hpx/issues).  Future major releases of HPX-5 will be delivered semi-annually although bug fixes will be made available between major releases.
-HPX-5 is developed through an agile process that includes continuous integration, regular point releases, and frequent regression tests for correctness and performance.  Users can submit issue reports to the development team through the HPX-5 web site.  Future major releases of HPX-5 will be delivered semi-annually although correctness and performance bug fixes will be made available between major releases.  Users may download official releases as well as the latest development branch of the code.
+The HPX-5 source code is released under the BSD open-source license and is distributed with a complete set of tests along with selected sample applications. HPX-5 is funded and supported by the DoD, DoE, and NSF, and is used actively in projects such as PSAAP, XPRESS, XSEDE. Further information and downloads for HPX-5 can be found at http://hpx.crest.iu.edu.
 == Quick Start Instructions ==
-If you plan to use HPX–5, we suggest to start with the latest released version (currently HPX–5 v1.1.0) which can be downloaded from https://hpx.crest.iu.edu/download.
+If you plan to use HPX–5, we suggest to start with the latest released version (currently HPX–5 v2.0.0) which can be downloaded from https://hpx.crest.iu.edu/download.
 Follow the installation directions located under hpx/INSTALL