Runtime Systems

Sonia Sachs held the first Runtime Systems Summit on April 9, 2014. The contents of the meeting are below, revised with updates by the attendees.

Exascale Runtime Systems Summit Plan and Outcomes

Summit Goals

Discuss current challenges in Exascale runtime systems: we would like to create an articulation of these challenges that is commonly accepted by the community. We want to discuss how to leverage recent reports and community input on challenges that we face in the area of runtime systems to create a complete set of challenges and discuss the current state-of-the-art to deal with them. Develop a set of questions that must be answered in this area. An initial set of questions is included in this document. Generate a roadmap for generating a unified runtime systems architecture for Exascale systems that has broad community acceptance and that leverages investments: Not a unified runtime software for Exascale To be clear: this summit is not an opportunity for participants to promote their current research agenda, but to take a fresh look into the future needs for Exascale runtime systems. The goal for this summit is to develop a unified runtime architecture with common components, not a single unified runtime software.

Summit rule: Participants should not promote their current research agenda

• Generate a roadmap for achieving a unified runtime systems architecture for Exascale systems
  • Reach consensus on the top six challenges and solutions for them.
  • Agree on a comprehensive set of questions that must be answered in order to achieve such architecture
  • Current known answers to posed questions
• Generate a roadmap for a research program on runtime systems
  • Consistent with achieving a unified runtime systems architecture
  • Discuss future workshop
• Prepare for writing a report

Plan to create the Unified Runtime Systems Architecture Roadmap

We need to leverage current investments in runtime systems: OCR, HPX, ARTS, SEEC, GVR runtime and runtimes to support advance/extended MPI and Global Arrays

• Agree on top six (6) challenges and solutions (1 hour)
  • Strawman set of challenges: slide 3
• For each challenge: discuss current state-of-the-art and how is challenge addressed in existing runtime systems to be leveraged? (1-2 hours)
• Agree on a set of top questions to be answered (1 hours)
  • Strawman set of questions: slide 4
• For each question: discuss currently known answers and how existing runtime systems answer it? (1-2 hours)
• Vision (1-2 hours)
  • what are major components?
  • programming interfaces and interfaces to the OS
  • How do we measure success

Strawman set of challenges

For the different execution models, key abstractions need to be identified and jointly supported by the runtime system, compilers, and hardware architecture. A large number of lightweight tasks and their coordination will need runtime support that is capable of dealing with system heterogeneity and with end-to-end asynchrony. Locality-aware, dynamic task scheduling will need runtime support so that it is possible to continuously optimize when code or data should be moved. Task coordination/synchronization primitives that are best suited to support exascale systems need to be identified. Load imbalances created by a large number of sources for non-uniform execution rates will require runtime support to dynamic load balancing.

• Key abstractions need to be identified
  • and jointly supported by the runtime system, compilers, and hardware architecture.
• Runtime support for lightweight tasks and their coordination
  • capable of dealing with system heterogeneity and with end-to-end asynchrony.
• Runtime support for locality-aware, dynamic task scheduling
  • Enabling continuously optimizing code or data movement.
• Need for task coordination and synchronization primitives
• Runtime support for dynamic load balancing
  • To deal with load imbalances created by a large number of sources for non-uniform execution rates

• What is the current known key abstractions? Which key abstractions are currently supported by runtime systems to be leveraged?
• For each of these challenges: What is the current state-of-the-art on such runtime support? How is this done in runtime systems to be leveraged?

Strawman Set of Questions

• Runtime system software architecture
  • What would the principal components be? What are the semantics of these components? What is the role of the different execution models?
  • What are the mechanisms for managing processes/threads/tasks and data?
  • What policies and/or mechanisms will your runtime use to schedule code and place data?
  • How does the runtime dynamically adapt the schedule and placement so that metrics of code-data affinity, power consumption, migration cost and resiliency are improved?
  • How does the runtime manage resources (compute, memory, power, bandwidth) to meet a power, energy and performance objective?
  • How does the runtime scale?
  • What is the role of a global address space or a global name space?
  • What programming models will be supported by the runtime architecture?
  • What OS support should be assumed?
• Community buy-in:
  • How do we achieve community buy-in to an envisioned runtime architecture and semantics?
  • We need a process to continuously evaluate refine the envisioned runtime architecture and semantics while keeping focus on achieving an Exascale runtime system.
  • What should this process be?

Current Runtime Investments

ASCR has made a number of investments in runtime system software research for Exascale. In the 2012 X-Stack program [2], projects include application-driven runtime systems support. Research in this area is concerned with maximizing concurrency efficiency, properly dealing with asynchrony of computation and communication, exploiting data locality, minimizing data movement, managing faults, and the needed support for heterogeneous computing elements, sound semantics for programmability, support for novel programming models, and delivery of an efficient execution environment to application developers. A number of runtime systems are currently being pursued: OCR, HPX, ARTS, SEEC, GVR runtime [2] and runtimes to support advance/extended MPI and Global Arrays [3]. Research of system-driven runtime systems, supported by the 2013 OS/R Program [4], is concerned with mechanisms, including semantics, of “common runtime services,” described in the OS/R report [5]. Examples of such mechanisms are thread management, low-level communication services, and resource management. Tight interaction among the different runtime service components has been identified as essential in order to deal with challenges of resilience, asynchronous computations, and locality of computation. A number of approaches of systems-driven runtime are currently being pursued in the ARGO, HOBBES, and X-ARCC projects. Research on applications-driven and systems-driven runtime systems are addressing challenges of resilience, power, hierarchical memory management, unprecedented parallelism, heterogeneity of hardware resources, locality and affinity management. DOE ASCR has insisted that focus on self-aware, dynamical systems should guide most of the research solutions in these two categories. DOE ASCR has also strongly recommended that close coordination be established among the various runtime system research projects. However, a community-wide involvement in defining runtime architecture for Exascale computing remains elusive.

• 2012 X-Stack program [2]: application-driven runtime systems support:
  • maximizing concurrency efficiency,
  • dealing with asynchrony of computation and communication,
  • exploiting data locality,
  • minimizing data movement,
  • managing faults,
  • support for heterogeneous computing elements,
  • semantics for programmability,
  • support for novel programming models
• Runtime systems to be leveraged
  • OCR, HPX, ARTS, SEEC, GVR runtime and runtimes to support advance/extended MPI and Global Arrays
• Mapping important questions to projects:
  • https://www.xstackwiki.com/index.php/Runtimes_(application-facing)

Remember to say that table of questions will be extended to include these three projects. Research of system-driven runtime systems, supported by the 2013 OS/R Program [4], is concerned with mechanisms, including semantics, of “common runtime services,” described in the OS/R report [5]. Examples of such mechanisms are thread management, low-level communication services, and resource management. Tight interaction among the different runtime service components has been identified as essential in order to deal with challenges of resilience, asynchronous computations, and locality of computation. A number of approaches of systems-driven runtime are currently being pursued in the ARGO, HOBBES, and X-ARCC projects. Research on applications-driven and systems-driven runtime systems are addressing challenges of resilience, power, hierarchical memory management, unprecedented parallelism, heterogeneity of hardware resources, locality and affinity management. DOE ASCR has insisted that focus on self-aware, dynamical systems should guide most of the research solutions in these two categories. DOE ASCR has also strongly recommended that close coordination be established among the various runtime system research projects. However, a community-wide involvement in defining runtime architecture for Exascale computing remains elusive.

• 2013 OS/R Program [4]: systems-driven mechanisms described in the OS/R report:
  • thread management,
  • low-level communication services,
  • resource management,
  • different runtime service components tightly connected to deal with challenges:
    • resilience,
    • asynchronous computations,
    • and locality of computation.
• Runtime systems approaches to be leveraged in ARGO, HOBBES, and X-ARCC projects.
• Mapping important questions to projects:
  • Not yet available
    • To be completed after upcoming OS/R semi-annual review

Summit Outcome: Top Challenges

• Growing gap between communication and computation
• Scalability & Starvation: dominant parameters to optimize, critical path management
• Locality and data movement: need terminology for inter and intra
• Power is critical
• Overhead
• Resilience: scalability and power problems exacerbates
• Load balancing: contention, hot spots,
• Heterogeneity: performance irregularities, static and dynamic, heterogeneity in storage/memory
• In-situ data analysis and mgmt: new dimension of interoperability
• Exploitation of runtime information (introspection), feedback control of performance data, managing performance data
• Resource allocation
• Scheduling & workflow orchestration
• Complexity/optimization/tuning
• Portability
• Synchronization: event-driven, mutual exclusion, barriers, phasers
• Computing everywhere
• Name space: both data and computation, includes location management
• Support tools
• Support for migratable computational units
• Hardware support, tight-coupling
• Expose some of runtime elements to system managers

Summit Outcome: Top Challenge Classes

We need to separate Problems from Solutions Runtime services: external view, tuning knobs, quality of service metrics (locality is a prime one) Internal runtime architecture: question on global optimization of runtime services. Need to time and energy optimize. Runtime systems serves one application, with instrospection. OS understands the system and the workload.

1. Locality and data movement: need terminology for inter and intra, dynamic decisions, handling variability, conflict of optimizing locality, data movement and costs of dynamic scheduling- questions of policy. Synchronization: event-driven, mutual exclusion, barriers, phasers. Overhead. Growing gap between communication and computation
2. Resilience: scalability and power problems exacerbates.
3. Variability. Static and Dynamic. Power management. Load balancing: contention, hot spots. Exploitation of runtime information (introspection), feedback control of perfomance data, managing performance data
4. Heterogeneity: performance irregularities, static and dynamic, heterogeneity in storage/memory. Computing everywhere.
5. Scalability & Starvation: dominant parameters to optimize, critical path management. Name space: both data and computation, includes location management. Complexity/optimization/tuning
6. Portability and interoperability. In-situ data analysis and mgmt: new dimension of interoperability: runtime systems composability.
7. Resource allocation. Scheduling & workflow orchestration. Cross jobs (apps) scheduling: OS role. Focus scheduling for one job. Support for migratable computational units. Hardware support, tight-coupling. Expose some of runtime elements to system managers
8. Usability. Support tools.

Summit Outcome: Challenge Problems

• For each challenge problem, we want to give examples in the context of challenge problems
  • Vivek suggested one multi-physics challenge problem.

HOMEWORK - ALL

• Identify and describe challenge problems

Summit Outcome: Key Abstractions

• Unit of computation
  • attributes: locality, synchronization, resilience, critical path
• Naming: data, computation, objects that combine both (active objects)
• Global side-effects: programming model abstraction?
• Execution Model
• Machine Model, Resources: memory, computation, storage, network, …
• Locality and affinity, hierarchy
• Control State: collective of info distributed across the global system that determines the next state of the machine. Distributed snapshot of the system. Logical abstraction, how to reason about the system.
• Enclave
• Scheduler: local scheduler of a single execution stream
• Execution Stream: something that has hardware associated with
• Communication data transfer
• Concurrency patterns, synchronization
• Resilience, detection, fault model

Summit Outcome: Runtime Services

Runtime Services

Runtime services: external view, tuning knobs, quality of service metrics (locality is a prime one)

• Schedule and execute threads/tasks/work unit, including code generation
• Resource allocation (give me resources dynamically, as needed, release resources): including networks. heterogeneity
• Introspection services: info about power, performance, heterogeneity. Variability.
• Creation, translation, isolation, security, release: name space, virtualization
• Communication of data and code, including synchronization (event-oriented). Migration services. Move work, move data. Is not separate from the communication services, it is composed with.
• Concurrency control: isolation, atomics (it gets into scheduling?)
• Location and Affinity/Locality services: map to some things that are mentioned above. Provides information and does binding.
• Express error checking/detection and recovery. Allows to specify resilience properties. Both to computation and data and hardware resources.
• Load balancing. Scheduling.
• OS requests services from the runtime: give me back resources that I gave you, tell runtime to graceful degradation/shutdown
• Services can make requests to other services, e.g., tools

Service Attributes

• How the service will be provided?
• Expected resilience
• Expected resources usage
• Persistence of memory
• Locality attributes

Homework - Wilf, Vivek, Kathy, Vijay

For the different execution models, key abstractions need to be identified and jointly supported by the runtime system, compilers, and hardware architecture. A large number of lightweight tasks and their coordination will need runtime support that is capable of dealing with system heterogeneity and with end-to-end asynchrony. Locality-aware, dynamic task scheduling will need runtime support so that it is possible to continuously optimize when code or data should be moved. Task coordination/synchronization primitives that are best suited to support exascale systems need to be identified. Load imbalances created by a large number of sources for non-uniform execution rates will require runtime support to dynamic load balancing.

• Refine key abstractions and their definition
• Using refined key abstractions, define runtime services
• Create a matrix for mapping on what current investments (including Charm++ runtime) provide these services and a brief description on how the services are provided.

Deadlines

• Initial draft to be distributed to summit participants: April 22
• Final draft including comments/suggestions from summit participants: April 30

Summit Outcome: Community Buy-In

• Presentation of solutions by Wilf (slides 15- 20)
• Discussions of the presented ideas
• More questions than we had time for:
  • We will post Wilf’s presentation in the xstack wiki
  • Wilf will present these again at the X-Stack PI meeting
  • Summit participants are encouraged to send Wilf and I comments/questions/suggestions
  • We will encourage X-Stack meeting participants to give us comments/questions/suggestions
• Ecosystem Creation: How do we achieve community buy-in to an envisioned runtime architecture and semantics?
• Process: We need a process to continuously evaluate refine the envisioned runtime architecture and semantics while keeping focus on achieving an Exascale runtime system. What should this process be?

Ecosystem Creation

• Establish an open, transparent environment where the solution is not pre-determined
• Provide an organic process for community decision-making, ensuring that the best solution wins
• Avoid a single player or clique dominating
• Lower the barrier to participation by providing stable, reliable releases of candidate solutions to a broad audience

Process

• Build an independent, open-source foundation that ensures the different projects can be continuously available, evolved, and supported.
• The different projects will evolve based on the contributions made. As solutions demonstrate their superiority, they will attract more contributions as well as consensus.
• The community will organically migrate to the superior solution.
• DOE can continuously view progress and help fund projects to cover any critical shortfalls.

Who is in the community?

1. Exascale Computing Research Community (us)
2. High Performance Computing User Community (current users)
3. Academic Community (future users)
4. Application Development Community (scientists and engineers)
5. Software Development Community
6. Hardware Vendors

Community Services

• Project Team Infrastructure - e.g. source code control, tooling, debuggers, collaboration/communication
• Release Engineering
• Technical Support
• IP management
• Education, instruction and training
• Community Development

Build on Experience: Community 2.0

Learn from the best Eclipse, Apache, Mozilla

• Building the community/ecosystem is top priority
• Support multiple projects and give them autonomy
• Support Regular Community Interaction
• Long-term commitment to quality through education and process

Avoid the pitfalls

• Commercial control of the purse strings leads to community breakdown
• Get to community support quickly and maintain community control

== Summit Outcomes

Comprehensive Set of Questions

Runtime system software architecture

• What are the major services provided by this architecture?
• What is the strategy that the runtime system has to embody? What is the role of the different execution models?
• What would the principal components be? What are the semantics of these components?
• What are the mechanisms for managing and scheduling units of computation and data?
• How does the runtime dynamically adapt the schedule and placement so that metrics of code-data affinity, power consumption, migration cost and resiliency are improved?
• How does the runtime manage resources (compute, memory, power, bandwidth) to meet a power, energy and performance objective? How are resources exposed?
• How does the runtime scale? How does the runtime ensures its scalability?
• What is the role of name/address spaces? Are they global or not? What is their scope?
• What programming models will be supported by the runtime architecture?
• How will composability be enabled?
• What OS support is assumed? What can the OS ask/expect from runtimes?
• What can compilers ask/expect of runtimes? What runtimes can ask/expect of compilers?
• Just in time compilation: what is the runtime support needed?
• What hardware support is assumed, can be exploited, can helpful? What is the machine model assumed?
• What is the cost model assumed (energy, performance, resilience)?
• How does the runtime system enables use of application or system info for resilience? In general, how does the runtime system uses information?
• What tools expect from runtimes, and what runtimes expect from tools?

Runtime Systems Major Components

• Unit of computation manager and scheduling
• Name service for everything that one wants to virtualize. Address allocation/translation.
• Data distribution and redistribution
• Locality management
• Power management
• Communication interfaces and/or infrastructure.
• Network I/O
• Active storage (compute in storage). I/O, locality
• Load balancers
• Location managers
• Prefetcher for explicit memory mgmt
• Lightweight migratable threads
• Introspection management. Monitoring/tools interface
• Reliable data store. I/O is embedded here.
• Global termination detection
• Event and synchronization framework
• Failure detection
• Failure recovery
• Adaptive controller
• Interoperability (with in-situ analysis, visualization, etc.)
• Composability manager

Runtime Systems vision

• Enable efficient exploitation of exascale hardware resource by applications.
• Addressing the identified exascale challenges
• Exploiting runtime information not available to compilers or programs
• Responsive to asynchrony
• Capable of delivering the services identified.

= Homework - Kathy

• Send us the mission statement, with input from; Sanjay, Vivek, Thomas, Costin, Ron (composability), Vijay, Milind