Date: 2019-06-03
Audience: SG1, SG14, LEWG
Authors: Gordon Brown, Ruyman Reyes, Michael Wong, Mark Hoemmen, Jeff Hammond, Tom Scogland
Emails: gordon@codeplay.com, ruyman@codeplay.com, michael@codeplay.com, mhoemme@sandia.gov, jeff.science@gmail.com, tscogland@llnl.gov
Reply to: gordon@codeplay.com
This paper is the result of discussions from man contributors within the heterogeneous C++ group, including H. Carter Edwards, Thomas Rodgers, Patrice Roy, Carl Cook, Jeff Hammond, Hartmut Kaiser, Christian Trott, Paul Blinzer, Alex Voicu, Nat Goodspeed and Tony Tye.
For the earlier changelogs from prior to the split from P0796 see Appendix A.
This paper is the result of a request from SG1 at the 2018 San Diego meeting to split [17] into two separate papers, one for the high-level interface and one for the low-level interface. This paper focusses on the low-level interface; a mechanism for discovering the topology and affinity properties of a given system. [18] focusses on the high-level interface, a series of properties for querying affinity relationships and requesting affinity on work being executed.
Computer systems are no longer homogeneous platforms. From desktop workstations to high-performance supercomputers, and from mobile devices to purpose-built embedded SoCs, every system has some form of co-processor along side the traditional multi-core CPU, and often more than one. Furthermore, the architectures of these co-processors range from many-core CPUs, GPUs, FPGAs and DSPs to specifically designed vision and machine learning processors. In larger supercomputer systems there are thousands of these processors in some configuration of nodes, connected physically or via network adapters.
The way these processors access memory is also far from homogeneous. For example, the system may present a single shared virtual address space [21] [22], or it may have different address spaces mutually inaccessible other than through special functions [4]. Different memory regions may have different levels of consistency, cache coherency, and support for atomic operations. Different parts of the system may have different access latencies or bandwidths to different memory regions (so-called “NUMA affinity regions”) [2]. Some parts of memory may be persistent. Different systems may configure the same types of memory in different ways around the processors.
In order to program these new systems and the architectures that inhabit them, it’s vital that applications are capable of understating both what architectures are available and the properties of those architectures, namely their observable behaviors, capabilities and limitations. However, the current C++ standard provides no way to achieve this, so developers have to rely entirely on third party and operating system libraries.
This paper seeks to define, within C++, a facility for discovering execution resources available to a system that are capable of executing work, and for querying their properties.
However, it is not the goal of this proposal to introduce support in the C++ language or the standard library for all of the various heterogeneous architectures available today. The authors of this paper recognize that this is unrealistic as it would require significant changes to the C++ machine model and would be extremely volatile to future developments in architecture and system design.
Instead, it seeks to define a single, unified, and stable layer in the C++ Standard Library. Applications, libraries, and programming models (such as SYCL [3], Kokkos [19], HPX [13] or TBB [12]) can build on this layer; hardware vendors can support it via standards such as OpenCL [4], CUDA [20], OpenMP [6], MPI [16], Hwloc [2], HSA [5] and HMM [21]; and it can be extended when necessary.
This layer will not be characterized in terms of specific categories of hardware such as CPUs, GPUs and FPGAs as these are broad concepts that are subject to change over time and have no foundation in the C++ machine model. It will instead define a number of abstract properties of system architectures that are not tied to any specific hardward.
The initial set of properties that this paper would propose be defined in the C++ standard library would reflect a generalization of the observable behaviors, capabilities and limitations of common architectures available in heterogeneous and distributed systems today. However the intention is that the interface be extensible so that that vendors can provide their own extensions to provide visibility into the more niche characteristics of certain architectures.
It is intended that this layer be defined as a natural extension of the Executors proposal, a unified interface for execution. The current executors proposal [[14]][p0443] already provides a route to supporting heterogeneous and distributed systems, however it is missing a way to identify what architectures a system has.
There are many reasons why such a feature within C++ would benefit developers and the C++ ecosystem as a whole, and those can differ from one domain to another. We’ve attempted to outline some of these benefits here.
The clearest benefit is performance. Exposing, even at an abstract level, the properties of the underlying architecture that a program is running on, allows application and libraries to be fine tuned. This may result in significant performance improvements that would only otherwise be possible via third party or operating system libraries [1] [7] [8] [9] [10] [11] [15].
This includes but is not limited to how to structure data to ensure access patterns along with execution on the architecture to achieve coalesced memory access and optimal cache utilization and where to initialize data to make efficient use of hardware locality and process affinity.
There is a general trend to move towards a unified address space in heterogeneous and distributed systems via standards like HMM. However, there are still many architectures that still require distinct address spaces, are not yet in a position to move to a single address space, and may never be. Even if you were to consider a single unified address the ultimate goal for heterogeneous and distributed systems, this actually makes the case for affinity in C++ stronger. As long as different address spaces exist, the distinction between different hardware memory regions and their capabilities is clear, but with a single unified address space, potentially with cache coherency, distinguishing different memory regions becomes much more subtle. Therefore, it becomes much more important to understand the various memory regions and their affinity relationships in order to achieve good performance on various architectures.
C++ is a major language when it comes to heterogeneous and distributed computing, and while it is a rapidly growing domain, it is still very challenging to develop in. There are a large number of C++ based third party and OS libraries. However, developing for heterogeneous and distributed systems often involves a combination of these libraries, which introduces a number of challenges
Firstly it’s common that different architectures are discovered via different libraries, You may want to use CUDA for NVidia GPUs, OpenMP for Intel CPUs, SYCL for Intel GPUs, Hwloc for the higher-level nodes and so on. This means that you have to collect together resources discovered from a different libraries, which very often do not provide a consistent representation or any form of interoperability, and find some way for them to represent them in a coherent view.
Secondly, many of these libraries report the same underlying hardware. For example OpenMP, SYCL and Hwloc will all report the same Intel CPU. This means you have to collate the resources together such that resources from different libraries representing the same hardware are joined together, to avoid resource contention.
There are many architectures available within heterogeneous and distributed systems which cannot support the full range of C++ features. This includes, but is not limited to dynamic allocation, recursion, dynamic polymorphism, RTTI, double precision floating point and some forms of atomic operations.
It’s crucial to allow developers to identify these limitations and which apply to the architecture they are running on, because in many cases if a C++ feature that is not supported on the architecture is used, the application would fail to execute or potentially crash.
Developing algorithms for heterogeneous and distributed systems requires at least an abstract understanding of the underlying architecture(s) being targeted in order to achieve optimal performance. In some cases, it may call for a much more in-depth understanding. This means that each algorithm may require a different implementation on each architecture.
Another factor here is that in many heterogeneous and distributed programming models, the architectures available on a particular system are not known until runtime, where the topology of the system is discovered.
Having a unified interface for performing this topology discovery and querying the properties of the architectures available on a system would dramatically improve developers’ ability to write generic algorithms.
Providing support for heterogenous and distributed computing as a first-class citizen of C++ will improve its accessibility and increase its utilization in libraries and applications, ultimately making the ecosystem stronger. This will become increasingly more important as heterogeneous and distributed computing becomes crucial to gaining the necessary performance in applications in more domains of C++.
The C++ standard is in a crucial position for heterogeneous and distributed computing domains. It is the common point between a number of different programming languages, models and libraries targeting a wide range of different architectures. This means that C++ has a unique opportunity to provide a single standard that not only covers the requirements of a single domain, but all of them, allowing for a convergence within the ecosystem and much more interoperability across different architectures.
For example, a unified C++ interface for topology discovery could provide access to GPUs from Nvidia, AMD, Intel, and ARM via their respective open standards or proprietary frameworks. At the same time, it could give access to NUMA-aware systems via Hwloc.
Another example of this is that while Hwloc is highly used in many domains, it now does not always accurately represent existing systems. This is because Hwloc presents their topology as strictly hierarchical, which no longer accurately describes many systems. A unified C++ interface does not need to be bound to the limitations of a single library, and can provide a much broader representation of a system’s execution resource topology.
Below we outline a proposed direction:
Align with the direction of the unified executors proposal [[14]][p0443].
Propose an abstract definition of an execution resource, as a hardware or software abstraction capable of creating execution agents.
Propose an abstract definition of an execution resource topology, as a non-hierarchical, non-acyclic graph of execution resources with various different kinds of connections.
Propose an interface for inspecting a system’s execution resource topology from various different perspectives, including, but not limited to:
Propose an interface for querying properties of an execution resource, reflecting the observable behaviors, capabilities and limitations of the architecture the execution resource represents.
Propose an interface for querying the relative affinity properties between different execution resources and memory regions within a system’s execution resource topology.
Propose an interface for binding execution agents and initialization of data to specific execution resources.
As a result of the above this paper may also:
Would SG1 like to see a continued effort to pursue the goals outlined in this paper?
Does SG1 believe the proposed direction laid out in this paper is suitable to achieve those goals?
[1] The Design of OpenMP Thread Affinity
[2] Portable Hardware Locality
[3] SYCL 1.2.1
[4] OpenCL 2.2
[5] HSA
[6] OpenMP 5.0
[7] cpuaff
[8] MEMKIND
[9] Solaris pbind()
[10] Linux sched_setaffinity()
[11] Windows SetThreadAffinityMask()
[12] TBB
[13] HPX
[p0443]:
http://wg21.link/p0443
[[14]][p0443] A Unified Executors Proposal for C++
[15] Exposing the Locality of new Memory Hierarchies to HPC Applications
[16] MPI
[17] Supporting Heterogeneous & Distributed Computing Through Affinity
[18] Executor properties for affinity-based execution
[19] Kokkos project
[20] CUDA
[21] Heterogeneous Memory Management
[22] OpenCL 2.x Shared Virtual Memory
execution_resource.execution_resource: it now either consistently identifies some underlying resource, or is invalid; context creation rejects an invalid resource.sterthis_thread::bind & this_thread::unbind interfaces.execution_resources iterable by replacing execution_resource::resources with execution_resource::begin and execution_resource::end.size and operator[] for execution_resource.this_system::get_resources to this_system::discover_topology.memory_resource to represent the memory component of a system topology.can_place_memory and can_place_agents from the execution_resource as these are no longer required.memory_resource and allocator from the execution_context as these no longer make sense.affinity_query to be between an execution_resource and a memory_resource.this_thread::bind & this_thread::unbind for binding and unbinding a thread of execution to an execution resource.bulk_execution_affinity executor properties for specifying affinity binding patterns on bulk execution functions.