XPRESS

From Extreme Scale Software Stack
Jump to: navigation, search
XPRESS
XPRESS-Logos.png
Team Members SNL, IU, LBNL, LSU, ORNL, UNC/RENCI, UH, UO
PI Ron Brightwell (SNL)
Co-PIs Andrew Lumsdaine (IU), Hartmut Kaiser (LSU), Barbara Chapman (UH), Allen Maloney (UO), Chris Baker (ORNL), Allan Porterfield (UNC/RENCI), Alice Koniges (LBNL)
Website http://xstack.sandia.gov/xpress
Download {{{download}}}

eXascale Programming Environment and System Software or XPRESS


Contents

Team Members

XPRESS-Focus-Area.png

Goals, Objectives, and Approach

  • Goals:
    • Enable exascale performance capability for current and future DOE applications
    • Develop and deliver a practical computing system software X‐stack, “OpenX”, for future practical DOE computing systems
    • Provide programming methods, environments, languages, and tools for effective means of expressing application and system software for portable exascale system execution
  • Objectives:
    • Derive a dynamic adaptive introspective strategy for exploiting opportunities and addressing critical exascale technology challenges in the form of an abstract execution model
    • Devise a software architecture as a framework for future exascale system design and implementation
    • Implement core interrelated and interoperable components of the software architecture to realize a fully working and usable system
    • Test, evaluate, validate, and demonstrate correctness, performance, resiliency, and energy efficiency
    • Provide technology transfer through cooperative engagement of industry hardware and software vendors and national labs via documentation and training
  • Approach:
    • Research, develop, and deploy a software stack to exploit the ParalleX execution model

XPRESS-Chart.png


Software Stack

ParalleX Execution Model

  • An execution model to provide the governing principles of computation to guide the system co-design and interoperability of software component layers and portability across system classes
  • Goal is to provide conceptual foundation to dramatically increase efficiency and scalability through transition from static to dynamic resource management and task scheduling and exploitation of new sources of parallelism
  • Key semantic constructs
    • Active Global Address Space (“AGAS”) for single system image
    • First class lightweight user threads for medium­‐gain parallelism
    • “Parcels” message­‐driven computing for latency mitigation
    • Local Control Objects (“LCO”) for powerful system synchronization
  • Performance strategy
    • Scalability through lightweight thread level parallelism, overlapping successive phases of computation with powerful synchronization and elimination of global barriers, automatic exposure/exploitation of intrinsic meta‐data parallelism, effective use of finer grain parallelism through reduction of overhead that bounds granularity
    • Latency mitigation through parcels by reducing numbers of long distance actions (split­‐phase transactions), localizing remote data and work, migrating continuations to change locus of continued execution with data structure, lightweight thread context switching, and direct in­‐memory and parcel generation without thread instantiation, and locality semantics
    • Overhead reduction is derived by powerful semantics of synchronization for minimum work, optimized thread control
    • Contention amelioration through dynamic resource management


XPRESS-Software-Architecture.png

LXK: Lightweight eXascale Kernel

  • Based on Sandia’s Kitten lightweight kernel
  • Boots idenDcally to Linux
    • Repurposes basic Linux funcDonality (PCI, NUMA, ACPI, etc.)
    • Supports POSIX threads (NPTL) and OpenMP
    • Allows innovaDon in key areas
      • Memory management
      • Multicore messaging
      • Network stack optimizations
      • Fully tick‐less operation
    • 20K LOC
  • Allows for re‐thinking OS structure and implementation of
    • Lightweight asynchronous system services
    • Dynamic composability and modularity
    • Adaptive resource policy enforcement mechanisms
    • Interface to runtime system(s)
    • Integrated instrumentation and monitoring

Kitten Implementation

  • Monolithic, C code, GNU toolchain, Kbuild configuration
  • Supports x86­‐64 architecture only, considering port to ARM
    • Boots on standard PC architecture, Cray XT, and in virtual machines
    • Boots identically to Linux (Kitten bzImage and init_task)
  • Repurposes basic functionality from Linux
    • Hardware bootstrap
    • Basic OS kernel primitives (lists, locks, wait queues, etc.)
    • PCI, NUMA, ACPI, IOMMU, …
    • Directory structure similar to Linux, arch dependent/independent directories
  • Custom address space management and task management
    • User‐level API for managing physical memory, building virtual address spaces
    • User‐level API for creating tasks, which run in virtual address spaces
    • User‐level API for migrating tasks between cores

Kitten Thread Support

  • Kitten user‐applications link with standard GNU C library (Glibc) and other system libraries installed on the Linux build host
  • Functionality added to Kitten to support Glibc NPTL POSIX threads implementation
    • Futex() system call (fast user‐level locking)
    • Basic support for signals
    • Match Linux implementation of thread local storage
    • Support for multiple threads per CPU core, preemptively scheduled
  • Kitten supports runtimes that work on top of POSIX threads
    • Glibc GOMP OpenMP implementation
    • Sandia Qthreads
    • Probably others with a little effort

HPX Runtime System

XPRESS-HPX.png
  • A next‐generation runtime system software layer that supports the semantics of the ParalleX execution model for significant increase in efficiency and scalability
  • HPX‐3 provides
    • Existing early proof‐of‐concept software dynamic adaptive resource management, task scheduling, global name space, efficient powerful synchronization, and Parcel message­‐driven computation. Interfaces with conventional Unix­‐like OS.
    • HPX­‐4
    • Modular system software architecture with specified functionality, interfaces and protocols for intra­‐operability and interfaces to OS and programming environment
    • Introspection closed­‐loop system for
      • Resiliency through microcheckpointing
      • Dynamic load balancing
      • Power monitoring and control

XPI: Low‐Level Intermediate Form

  • User programming syntax and source‐to‐source compiler target for high‐level programming languages
  • Provides stable application plakorm based on HPX, which is expected to change underneath throughout project
  • A library of C­‐bindings to represent lowest­‐level semantics, policies, and mechanisms for ParalleX execution model
  • XPI construct classes
    • Process
    • Thread
    • Locality
    • Parcel
    • Future
    • Dataflow
    • Housekeeping

XPRESS Information Flow

  • Passing information between layers will be critical
    • In current systems information flows one direction
  • For Exascale static scheduling decisions will not work
    • Dynamic environment
      • Billion‐way parallelism
      • Resilience
      • Reliability
      • Energy
      • Shared resource contention
  • Feedback will be required

Performance Information as Glue

  • Performance informaDon
    • Current: post‐execution performance tools
    • Exascale: dynamic application introspection
  • For performance and reliability thread/core/node/system knowledge will be critical throughout the software stack
    • Interfaces designed to enable information flow
      • Utilities need to know current system performance
      • Utilities need to know how other utilities are reacting

Exascale Performance Observability

XPRESS-OpenX.png
  • Exascale requires a fundamentally new observability paradigm
    • Reflects translation of application and mapping of computation model to execution model
    • Designed specifically to support introspection of runtime performance for adaptation and control
    • Aware of multiple objectives
      • System‐level resource utilization data and analysis, energy consumption, and health information
  • Exascale observability abstraction
    • Inherent state of exascale execution is dynamic
    • Embodies non‐stationarity of performance, energy, resilience during application execution
    • Constantly shaped by the adaptation of resources to meet computational needs and optimize execution objectives

OpenX Software Stack and APEX

XPRESS-OpenX-Software-Stack.png

APEX

  • XPRESS performance measurement/analysis/introspection
    • Observation and runtime analysis of performance, energy, and reliability
    • Online introspection resource utilization, energy consumption, and health information
    • Coupling of introspection with OpenX software stack for self­‐adaptive control
  • APEX: Autonomic Performance Environment for eXascale
    • Support performance awareness and performance reactivity
    • Couple with application and execution knowledge
    • Serve top­‐down AND bottom‐up requirements of OpenX
    • Performance overlay on OpenX

APEX’s Role for Top­‐Down OpenX Requirements

  • Top-­‐down requirements are driven by:
    • Mapping of applications to the ParalleX model
    • Translation through the programming models and the language compilers into runtime operations and execution
    • Performance abstractions (PA) at each level define:
      • Set of parameters to be observed by the next levels down
      • Performance model to be evaluated and provide basis for control
    • Performance abstractions are coupled with observations through APEX’s hierarchical performance framework
    • Realization of control mechanisms (reactive to PA actualization)
  • Top­‐down view sees APEX functionality as part of the application’s execution, specialized with observability and introspection support built into each OpenX layer:
    • LXK OS – system resource, utilization, job contention, overhead
    • HPX – threads, queues, concurrency, remote, parcels, memory
    • XPI/Legacy – language‐level performance semantics

APEX’s Role for Bottom­‐Up OpenX Requirements

  • Bottom­‐up requirements are driven by:
    • Performance introspection across the OpenX layers
    • Enable dynamic, adaptive operation, and decision control
    • Online access and analysis of observations at different levels
    • Working model is multi‐parameter system optimization
  • APEX creates the performance feedback mechanisms and builds an efficient hierarchical infrastructure for connecting subscribers to runtime performance state
    • Intra‐level performance awareness for HPX and LXK
    • Interplay with the overall application dynamics (top­‐down)
      • Top­‐down requirement implement constraints
    • Performance information is the result of runtime analysis

Top­‐Down Development Approach

  • Define performance abstraction (focusing on higher level)
    • Specify observability requirements, and results semantics
    • Provide application context for association
    • What are the performance models, attributes, factors?
  • Build into legacy languages and XPI
    • APEX programming of PA infrastructure
    • Invokes HPX performance measurement API
  • Integrate TAU capabilities for APEX realization
    • Instrumentation
      • legacy programming (MPI, OpenMP, OpenACC)
      • imperative API (XPI) for ParalleX programming
    • TAU mapping for measurement contextualization
    • Wrapper interposition to intercept HPX runtime layer
  • Create introspection API for feedback


Legacy Application Migration and Interoperability

  • Seamless migration, no modification needed for apps
    • Retargeting OpenMP compiler to OpenX
    • Adapting MPI to OpenX, need MPI system‐level adjustment, and change of configuration logic
  • Two approaches for retargeting OpenMP via XPI:
    • Through a POSIX compliant interface using XPI, e.g. pthreads on top of XPI
      • +: No (or minor) modification needed to OpenMP compiler
      • -­: May only use a limited subset of OpenX features
    • Rewrite OpenMP compiler and runtime to XPI
      • +: Explore OpenX advanced features
  • OpenACC integration with OpenX
    • Adapt the OpenACC data movement mechanism to OpenX communication parcels in the AGAS
    • Integrate accelerator kernel execution of OpenACC with the OpenX execution model
  • Evaluation:
    • Applications: Mini­‐apps with MPI+OpenMP/OpenACC
    • Performance and productivity feedback to the OpenX implementation team
Personal tools
Namespaces

Variants
Actions
Navigation
Toolbox