Stanford University logo SLAC National Accelerator Laboratory logo Los Alamos National Laboratory logo NVIDIA logo Winner of the R&D 100 Award

Legion

A Data-Centric Parallel Programming System

Github

Publications

Table of Contents

Papers

Legion: Expressing Locality and Independence with Logical Regions PDF
Michael Bauer, Sean Treichler, Elliott Slaughter, Alex Aiken
In the International Conference for High Performance Computing, Networking, Storage and Analysis (SC 2012)
Abstract: Modern parallel architectures have both heterogeneous processors and deep, complex memory hierarchies. We present Legion, a programming model and runtime system for achieving high performance on these machines. Legion is organized around logical regions, which express both locality and independence of program data, and tasks, functions that perform computations on regions. We describe a runtime system that dynamically extracts parallelism from Legion programs, using a distributed, parallel scheduling algorithm that identifies both independent tasks and nested parallelism. Legion also enables explicit, programmer controlled movement of data through the memory hierarchy and placement of tasks based on locality information via a novel mapping interface. We evaluate our Legion implementation on three applications: fluid-flow on a regular grid, a three-level AMR code solving a heat diffusion equation, and a circuit simulation.

Language Support for Dynamic, Hierarchical Data Partitioning PDF
Sean Treichler, Michael Bauer, Alex Aiken
In Object Oriented Programming, Systems, Languages, and Applications (OOPSLA 2013)
Abstract: Applications written for distributed-memory parallel architectures must partition their data to enable parallel execution. As memory hierarchies become deeper, it is increasingly necessary that the data partitioning also be hierarchical to match. Current language proposals perform this hierarchical partitioning statically, which excludes many important applications where the appropriate partitioning is itself data dependent and so must be computed dynamically. We describe Legion, a region-based programming system, where each region may be partitioned into subregions. Partitions are computed dynamically and are fully programmable. The division of data need not be disjoint and subregions of a region may overlap, or alias one another. Computations use regions with certain privileges (e.g., expressing that a computation uses a region read-only) and data coherence (e.g., expressing that the computation need only be atomic with respect to other operations on the region), which can be controlled on a per-region (or subregion) basis.

We present the novel aspects of the Legion design, in particular the combination of static and dynamic checks used to enforce soundness. We give an extended example illustrating how Legion can express computations with dynamically determined relationships between computations and data partitions. We prove the soundness of Legion’s type system, and show Legion type checking improves performance by up to 71% by eliding provably safe memory checks. In particular, we show that the dynamic checks to detect aliasing at runtime at the region granularity have negligible overhead. We report results for three real-world applications running on distributed memory machines, achieving up to 62.5X speedup on 96 GPUs on the Keeneland supercomputer.

Realm: An Event-Based Low-Level Runtime for Distributed Memory Architectures PDF
Sean Treichler, Michael Bauer, Alex Aiken
In Parallel Architectures and Compilation Techniques (PACT 2014)
Abstract: We present Realm, an event-based runtime system for heterogeneous, distributed memory machines. Realm is fully asynchronous: all runtime actions are non-blocking. Realm supports spawning computations, moving data, and reservations, a novel synchronization primitive. Asynchrony is exposed via a light-weight event system capable of operating without central management.

We describe an implementation of Realm that relies on a novel generational event data structure for efficiently handling large numbers of events in a distributed address space. Micro-benchmark experiments show our implementation of Realm approaches the underlying hardware performance limits. We measure the performance of three real-world applications on the Keeneland supercomputer. Our results demonstrate that Realm confers considerable latency hiding to clients, attaining significant speedups over traditional bulk-synchronous and independently optimized MPI codes.

Structure Slicing: Extending Logical Regions with Fields PDF
Michael Bauer, Sean Treichler, Elliott Slaughter, Alex Aiken
In the International Conference for High Performance Computing, Networking, Storage and Analysis (SC 2014)
Abstract: Applications on modern supercomputers are increasingly limited by the cost of data movement, but mainstream programming systems have few abstractions for describing the structure of a program’s data. Consequently, the burden of managing data movement, placement, and layout currently falls primarily upon the programmer

To address this problem, we previously proposed a data model based on logical regions and described Legion, a programming system incorporating logical regions. In this paper, we present structure slicing, which incorporates fields into the logical region data model. We show that structure slicing enables Legion to automatically infer task parallelism from field non-interference, decouple the specification of data usage from layout, and reduce the overall amount of data moved. We demonstrate that structure slicing enables both strong and weak scaling of three Legion applications, including S3D, a production combustion simulation that uses logical regions and thousands of fields, with speedups of up to 3.68X over a vectorized CPU-only Fortran implementation and 1.88X over an independently hand-tuned OpenACC code.

Note: The following paper is a result of our collaboration with the ExaCT Combustion Co-Design Center and shows how a DSL compiler can be used to generate fast tasks for Legion applications.
Singe: Leveraging Warp Specialization for High Performance on GPUs PDF
Michael Bauer, Sean Treichler, Alex Aiken
In Principles and Practices of Parallel Programming (PPoPP 2014)
Abstract: We present Singe, a Domain Specific Language (DSL) compiler for combustion chemistry that leverages warp specialization to produce high performance code for GPUs. Instead of relying on traditional GPU programming models that emphasize data-parallel computations, warp specialization allows compilers like Singe to partition computations into sub-computations, which are then assigned to different warps within a thread block. Fine-grain synchronization between warps is performed efficiently in hardware using producer-consumer named barriers. Partitioning computations using warp specialization allows Singe to deal efficiently with the irregularity in both data access patterns and computation. Furthermore, warp-specialized partitioning of computations allows Singe to fit extremely large working sets into on-chip memories. Finally, we describe the architecture and general computation techniques necessary for constructing a warp-specializing compiler. We show that the warp-specialized code emitted by Singe is up to 3.75X faster than previously optimized data-parallel GPU kernels.

Note: The following paper is an example of a DSL compiler toolchain that targets Legion as a backend.
Exploring the Construction of a Domain-Aware Toolchain for High-Performance Computing PDF
Patrick McCormick, Christine Sweeney, Nick Moss, Dean Prichard, Samuel K. Gutierrez, Kei Davis, Jamaludin Mohd-Yusof
In the International Workshop on Domain-Specific Languages and High-Level Frameworks for High Performance Computing (WOLFHPC 2014)
Abstract: The push towards exascale computing has sparked a new set of explorations for providing new productive programming environments. While many efforts are focusing on the design and development of domain-specific languages (DSLs), few have addressed the need for providing a fully domain-aware toolchain. Without such domain awareness critical features for achieving acceptance and adoption, such as debugger support, pose a long-term risk to the overall success of the DSL approach. In this paper we explore the use of language extensions to design and implement the Scout DSL and a supporting toolchain infrastructure. We highlight how language features and the software design methodologies used within the toolchain play a significant role in providing a suitable environment for DSL development.

Regent: A High-Productivity Programming Language for HPC with Logical Regions PDF
Elliott Slaughter, Wonchan Lee, Sean Treichler, Michael Bauer, and Alex Aiken
In the International Conference for High Performance Computing, Networking, Storage and Analysis (SC 2015)
Abstract: We present Regent, a high-productivity programming language for high performance computing with logical regions. Regent users compose programs with tasks (functions eligible for parallel execution) and logical regions (hierarchical collections of structured objects). Regent programs appear to execute sequentially, require no explicit synchronization, and are trivially deadlock-free. Regent’s type system catches many common classes of mistakes and guarantees that a program with correct serial execution produces identical results on parallel and distributed machines.

We present an optimizing compiler for Regent that translates Regent programs into efficient implementations for Legion, an asynchronous task-based model. Regent employs several novel compiler optimizations to minimize the dynamic overhead of the runtime system and enable efficient operation. We evaluate Regent on three benchmark applications and demonstrate that Regent achieves performance comparable to hand-tuned Legion.

Dependent Partitioning PDF
Sean Treichler, Michael Bauer, Rahul Sharma, Elliott Slaughter, and Alex Aiken
In Object Oriented Programming, Systems, Languages, and Applications (OOPSLA 2016)
Abstract: A key problem in parallel programming is how data is partitioned: divided into subsets that can be operated on in parallel and, in distributed memory machines, spread across multiple address spaces.

We present a dependent partitioning framework that allows an application to concisely describe relationships between partitions. Applications first establish independent partitions, which may contain arbitrary subsets of application data, permitting the expression of arbitrary application-specific data distributions. Dependent partitions are then derived from these using the dependent partitioning operations provided by the framework. By directly capturing inter-partition relationships, our framework can soundly and precisely reason about programs to perform important program analyses crucial to ensuring correctness and achieving good performance. As an example of the reasoning made possible, we present a static analysis that discharges most consistency checks on partitioned data during compilation.

We describe an implementation of our framework within Regent, a language designed for the Legion programming model. The use of dependent partitioning constructs results in a 86-96% decrease in the lines of code required to describe the partitioning, the elimination of many of the expensive dynamic checks required for soundness by the current Regent partitioning implementation, and speeds up the computation of partitions by 2.6-12.7X even on a single thread. Furthermore, we show that a distributed implementation incorporated into the the Legion runtime system allows partitioning of data sets that are too large to fit on a single node and yields an additional 29X speedup of partitioning operations on 64 nodes.

Control Replication: Compiling Implicit Parallelism to Efficient SPMD with Logical Regions PDF
Elliott Slaughter, Wonchan Lee, Sean Treichler, Wen Zhang, Michael Bauer, Galen Shipman, Patrick McCormick and Alex Aiken
In the International Conference for High Performance Computing, Networking, Storage and Analysis (SC 2017)
Abstract: We present control replication, a technique for generating high-performance and scalable SPMD code from implicitly parallel programs. In contrast to traditional parallel programming models that require the programmer to explicitly manage threads and the communication and synchronization between them, implicitly parallel programs have sequential execution semantics and by their nature avoid the pitfalls of explicitly parallel programming. However, without optimizations to distribute control overhead, scalability is often poor.

Performance on distributed-memory machines is especially sensitive to communication and synchronization in the program, and thus optimizations for these machines require an intimate understanding of a program’s memory accesses. Control replication achieves particularly effective and predictable results by leveraging language support for first-class data partitioning in the source programming model. We evaluate an implementation of control replication for Regent and show that it achieves up to 99% parallel efficiency at 1024 nodes with absolute performance comparable to hand-written MPI(+X) codes.

Integrating External Resources with a Task-Based Programming Model PDF
Zhihao Jia, Sean Treichler, Galen Shipman, Michael Bauer, Noah Watkins, Carlos Maltzahn, Patrick McCormick and Alex Aiken
In the International Conference on High Performance Computing, Data, and Analytics (HiPC 2017)
Abstract: Accessing external resources (e.g., loading input data, checkpointing snapshots, and out-of-core processing) can have a significant impact on the performance of applications. However, no existing programming systems for high-performance computing directly manage and optimize external accesses. As a result, users must explicitly manage external accesses alongside their computation at the application level, which can result in both correctness and performance issues.

We address this limitation by introducing Iris, a task-based programming model with semantics for external resources. Iris allows applications to describe their access requirements to external resources and the relationship of those accesses to the computation. Iris incorporates external I/O into a deferred execution model, reschedules external I/O to overlap I/O with computation, and reduces external I/O when possible. We evaluate Iris on three microbenchmarks representative of important workloads in HPC and a full combustion simulation, S3D. We demonstrate that the Iris implementation of S3D reduces the external I/O overhead by up to 20x, compared to the Legion and the Fortran implementations.

In Situ Visualization with Task-based Parallelism PDF
Alan Heirich, Elliott Slaughter, Manolis Papadakis, Wonchan Lee, Tim Biedert and Alex Aiken
In the Workshop on In Situ Infrastructures for Enabling Extreme-scale Analysis and Visualization (ISAV 2017)
Abstract: This short paper describes an experimental prototype of in situ visualization in a task-based parallel programming framework. A set of reusable visualization tasks were composed with an existing simulation. The visualization tasks include a local OpenGL renderer, a parallel image compositor, and a display task. These tasks were added to an existing fluid-particle-radiation simulation and weak scaling tests were run on up to 512 nodes of the Piz Daint supercomputer. Benchmarks showed that the visualization components scaled and did not reduce the simulation throughput. The compositor latency increased logarithmically with increasing node count.

S3D-Legion: An Exascale Software for Direct Numerical Simulation of Turbulent Combustion with Complex Multicomponent Chemistry PDF
Sean Treichler, Michael Bauer, Ankit Bhagatwala, Giulio Borghesi, Ramanan Sankaran, Hemanth Kolla, Patrick S. McCormick, Elliott Slaughter, Wonchan Lee, Alex Aiken and Jacqueline Chen
In Exascale Scientific Applications: Scalability and Performance Portability (CRC Press, 2017)

A Distributed Multi-GPU System for Fast Graph Processing PDF, Software Release
Zhihao Jia, Yongkee Kwon, Galen Shipman, Pat McCormick, Mattan Erez and Alex Aiken
In the International Conference on Very Large Data Bases (VLDB 2018)
Abstract: We present Lux, a distributed multi-GPU system that achieves fast graph processing by exploiting the aggregate memory bandwidth of multiple GPUs and taking advantage of locality in the memory hierarchy of multi-GPU clusters. Lux provides two execution models that optimize algorithmic efficiency and enable important GPU optimizations, respectively. Lux also uses a novel dynamic load balancing strategy that is cheap and achieves good load balance across GPUs. In addition, we present a performance model that quantitatively predicts the execution times and automatically selects the runtime configurations for Lux applications. Experiments show that Lux achieves up to 20× speedup over state-of-the-art shared memory systems and up to two orders of magnitude speedup over distributed systems.

Dynamic Tracing: Memoization of Task Graphs for Dynamic Task-Based Runtimes PDF
Wonchan Lee, Elliott Slaughter, Michael Bauer, Sean Treichler, Todd Warszawski, Michael Garland and Alex Aiken
In the International Conference for High Performance Computing, Networking, Storage and Analysis (SC 2018)
Abstract: Many recent programming systems for both supercomputing and data center workloads generate task graphs to express computations that run on parallel and distributed machines. Due to the overhead associated with constructing these graphs the dependence analysis that generates them is often statically computed and memoized, and the resulting graph executed repeatedly at runtime. However, many applications require a dynamic dependence analysis due to data dependent behavior, but there are new challenges in capturing and re-executing task graphs at runtime. In this work, we introduce dynamic tracing, a technique to capture a dynamic dependence analysis of a trace that generates a task graph, and replay it. We show that an implementation of dynamic tracing improves strong scaling by an average of 4.9X and up to 7.0X on a suite of already optimized benchmarks.

Correctness of Dynamic Dependence Analysis for Implicitly Parallel Tasking Systems PDF
Wonchan Lee, George Stelle, Patrick McCormick and Alex Aiken
In the International Workshop on Software Correctness for HPC Applications (Correctness 2018)
Abstract: In this paper, we rigorously verify the correctness of dynamic dependence analysis, a key algorithm for parallelizing programs in implicitly parallel tasking systems. A dynamic dependence analysis of a program results in a task graph, a DAG of tasks constraining the order of task execution. Because a program is automatically parallelized based on its task graph, the analysis algorithm must generate a graph with all the dependencies that are necessary to preserve the program’s original semantics for any non-deterministic parallel execution of tasks. However, this correctness is not straightforward to verify as implicitly parallel tasking systems often use an optimized dependence analysis algorithm. To study the correctness of dynamic dependence analysis in a realistic setting, we design a model algorithm that captures the essence of realistic analysis algorithms. We prove that this algorithm constructs task graphs that soundly and completely express correct parallel executions of programs. We also show that the generated task graph is the most succinct one for a program when the program satisfies certain conditions.

Soleil-X: Turbulence, Particles, and Radiation in the Regent Programming Language PDF
Hilario Torres, Manolis Papadakis, Lluís Jofre, Wonchan Lee, Alex Aiken and Gianluca Iaccarino
In Bulletin of the American Physical Society (2018)
Abstract: The Predictive Science Academic Alliance Program (PSAAP) II at Stanford University is developing an Exascale-ready multi-physics solver to investigate particle-laden turbulent flows in a radiation environment for solar energy receiver applications. To simulate the proposed concentrated particle-based receiver design three distinct but coupled physical phenomena must be modeled: fluid flows, Lagrangian particle dynamics, and the transport of thermal radiation. Therefore, three different physics solvers (fluid, particles, and radiation) must run concurrently with significant cross-communication in an integrated multi-physics simulation. However, each solver uses substantially different algorithms and data access patterns. Coordinating the overall data communication, computational load balancing, and scaling these different physics solvers together on modern massively parallel, heterogeneous high performance computing systems presents several major challenges. We have adopted the Legion programming system, via the Regent programming language, and its task parallel programming model to address these challenges. Our multi-physics solver Soleil-X is written entirely in the high level Regent programming language and is one of the largest and most complex applications written in Regent to date. At this workshop we will give an overview of the software architecture of Soleil-X as well as discuss how our multi-physics solver was designed to use the task parallel programming model provided by Legion. We will also discuss the developmentexperience, scaling, performance, portability, and multi-physics simulation results.

Legate NumPy: Accelerated and Distributed Array Computing PDF
Michael Bauer and Michael Garland
In the International Conference for High Performance Computing, Networking, Storage and Analysis (SC 2019)
Abstract: NumPy is a popular Python library used for performing array-based numerical computations. The canonical implementation of NumPy used by most programmers runs on a single CPU core and only a few operations are parallelized across cores. This restriction to single-node CPU-only execution limits both the size of data that can be processed and the speed with which problems can be solved. In this paper we introduce Legate, a programming system that transparently accelerates and distributes NumPy programs to machines of any scale and capability typically by changing a single module import statement. Legate achieves this by translating the NumPy application interface into the Legion programming model and leveraging the performance and scalability of the Legion runtime. We demonstrate that Legate can achieve state-of-the-art scalability when running NumPy programs on machines with up to 1280 CPU cores and 256 GPUs, allowing users to prototype on their desktop and immediately scale up to significantly larger machines. Furthermore, we demonstrate that Legate can achieve between one and two orders of magnitude better performance than the popular Python library Dask when running comparable programs at scale.

A Constraint-Based Approach to Automatic Data Partitioning for Distributed Memory Execution PDF
Wonchan Lee, Manolis Papadakis, Elliott Slaughter and Alex Aiken
In the International Conference for High Performance Computing, Networking, Storage and Analysis (SC 2019)
Abstract: Although data partitioning is required to enable parallelism on distributed memory systems, data partitions are not first class objects in most distributed programming models. As a result, automatic parallelizers and application writers encode a particular partitioning strategy in the parallelized program, leading to a program not easily configured or composed with other parallel programs.

We present a constraint-based approach to automatic data partitioning. By introducing abstractions for first-class data partitions, we express a space of correct partitioning strategies. Candidate partitions are characterized by partitioning constraints, which can be automatically inferred from data accesses in parallelizable loops. Constraints can be satisfied by synthesized partitioning code or user-provided partitions. We demonstrate that programs auto-parallelized in our approach are easily composed with manually parallelized parts and have scalability comparable to hand-optimized counterparts.

Pygion: Flexible, Scalable Task-Based Parallelism with Python PDF
Elliott Slaughter and Alex Aiken
In the Parallel Applications Workshop, Alternatives To MPI+X (PAW-ATM 2019)
Abstract: Dynamic languages provide the flexibility needed to implement expressive support for task-based parallel programming constructs. We present Pygion, a Python interface for the Legion task-based programming system, and show that it can provide features comparable to Regent, a statically typed programming language with dedicated support for the Legion programming model. Furthermore, we show that the dynamic nature of Python permits the implementation of several key optimizations (index launches, futures, mapping) currently implemented in the Regent compiler. Together these features enable Pygion code that is comparable in expressiveness but more flexible than Regent, and substantially more concise, less error prone, and easier to use than C++ Legion code. Pygion is designed to interoperate with Regent and can use Regent to generate high-performance CPU and GPU kernel implementations. We show that, in combination with high-performance kernels written in Regent, Pygion is able to achieve efficient, scalable execution on up to 512 nodes of the heterogeneous supercomputer Piz Daint.

HTR Solver: An Open-Source Exascale-Oriented Task-Based Multi-GPU High-Order Code for Hypersonic Aerothermodynamics PDF
Mario Di Renzo, Lin Fu and Javier Urzay
In Computer Physics Communications (2020)
Abstract: In this study, the open-source Hypersonics Task-based Research (HTR) solver for hypersonic aerothermodynamics is described. The physical formulation of the code includes thermochemical effects induced by high temperatures (vibrational excitation and chemical dissociation). The HTR solver uses high-order TENO-based spatial discretization on structured grids and efficient time integrators for stiff systems, is highly scalable in GPU-based supercomputers as a result of its implementation in the Regent/Legion stack, and is designed for direct numerical simulations of canonical hypersonic flows at high Reynolds numbers. The performance of the HTR solver is tested with benchmark cases including inviscid vortex advection, low- and high-speed laminar boundary layers, inviscid one-dimensional compressible flows in shock tubes, supersonic turbulent channel flows, and hypersonic transitional boundary layers of both calorically perfect gases and dissociating air.

Task Bench: A Parameterized Benchmark for Evaluating Parallel Runtime Performance PDF
Elliott Slaughter, Wei Wu, Yuankun Fu, Legend Brandenburg, Nicolai Garcia, Wilhem Kautz, Emily Marx, Kaleb S. Morris, Qinglei Cao, George Bosilca, Seema Mirchandaney, Wonchan Lee, Sean Treichler, Patrick McCormick, and Alex Aiken
In the International Conference for High Performance Computing, Networking, Storage and Analysis (SC 2020)
Abstract: We present Task Bench, a parameterized benchmark designed to explore the performance of distributed programming systems under a variety of application scenarios. Task Bench dramatically lowers the barrier to benchmarking and comparing multiple programming systems by making the implementation for a given system orthogonal to the benchmarks themselves: every benchmark constructed with Task Bench runs on every Task Bench implementation. Furthermore, Task Bench’s parameterization enables a wide variety of benchmark scenarios that distill the key characteristics of larger applications.

To assess the effectiveness and overheads of the tested systems, we introduce a novel metric, minimum effective task granularity (METG). We conduct a comprehensive study with 15 programming systems on up to 256 Haswell nodes of the Cori supercomputer. Running at scale, 100μs-long tasks are the finest granularity that any system runs efficiently with current technologies. We also study each system’s scalability, ability to hide communication and mitigate load imbalance.

An Implicitly Parallel Meshfree Solver in Regent PDF
Rupanshu Soi, Nischay Ram Mamidi, Elliott Slaughter, Kumar Prasun, Anil Nemili, and S.M. Deshpande
In the Parallel Applications Workshop, Alternatives to MPI+X (PAW-ATM 2020)
Abstract: This paper presents the development of a Regent based implicitly parallel meshfree solver for inviscid compressible fluid flows. The meshfree solver is based on the Least Squares Kinetic Upwind Method (LSKUM). The performance of the Regent parallel solver is assessed by comparing with the explicitly parallel versions of the same solver written in Fortran 90 and Julia. The Fortran code uses MPI with PETSc libraries, while the Julia code uses an MPI + X alternative parallel library. Numerical results are shown to assess the performance of these solvers on single and multiple CPU nodes.

Scaling Implicit Parallelism via Dynamic Control Replication PDF
Michael Bauer, Wonchan Lee, Elliott Slaughter, Zhihao Jia, Mario Di Renzo, Manolis Papadakis, Galen Shipman, Patrick McCormick, Michael Garland, and Alex Aiken
In Principles and Practices of Parallel Programming (PPoPP 2021)
Abstract: We present dynamic control replication, a run-time program analysis that enables scalable execution of implicitly parallel programs on large machines through a distributed and efficient dynamic dependence analysis. Dynamic control replication distributes dependence analysis by executing multiple copies of an implicitly parallel program while ensuring that they still collectively behave as a single execution. By distributing and parallelizing the dependence analysis, dynamic control replication supports efficient, on-the-fly computation of dependences for programs with arbitrary control flow at scale. We describe an asymptotically scalable algorithm for implementing dynamic control replication that maintains the sequential semantics of implicitly parallel programs.

An implementation of dynamic control replication in the Legion runtime delivers the same programmer productivity as writing in other implicitly parallel programming models, such as Dask or TensorFlow, while providing better performance (11.4X and 14.9X respectively in our experiments), and scalability to hundreds of nodes. We also show that dynamic control replication provides good absolute performance and scaling for HPC applications, competitive in many cases with explicitly parallel programming systems.

Index Launches: Scalable, Flexible Representation of Parallel Task Groups PDF
Rupanshu Soi, Michael Bauer, Sean Treichler, Manolis Papadakis, Wonchan Lee, Patrick McCormick, Alex Aiken, and Elliott Slaughter
In the International Conference for High Performance Computing, Networking, Storage and Analysis (SC 2021)
Abstract: It’s common to see specialized language constructs in modern task-based programming systems for reasoning about groups of independent tasks intended for parallel execution. However, most systems use an ad-hoc representation that limits expressiveness and often overfits for a given application domain. We introduce index launches, a scalable and flexible representation of a group of tasks. Index launches use a flexible mechanism to indicate the data required for a given task, allowing them to be used for a much broader set of use cases while maintaining an efficient representation. We present a hybrid design for index launches, involving static and dynamic program analyses, along with a characterization of how they’re used in Legion and Regent, and show how they generalize constructs found in other task-based systems. Finally, we present results of scaling experiments which demonstrate that index launches are crucial for the efficient distributed execution of several scientific codes in Regent.

DISTAL: The Distributed Tensor Algebra Compiler PDF
Rohan Yadav, Alex Aiken and Fredrik Kjolstad
In the Conference on Programming Language Design and Implementation (PLDI 2022)
Abstract: We introduce DISTAL, a compiler for dense tensor algebra that targets modern distributed and heterogeneous systems. DISTAL lets users independently describe how tensors and computation map onto target machines through separate format and scheduling languages. The combination of choices for data and computation distribution creates a large design space that includes many algorithms from both the past (e.g., Cannon’s algorithm) and the present (e.g., COSMA). DISTAL compiles a tensor algebra domain specific language to a distributed task-based runtime system and supports nodes with multi-core CPUs and multiple GPUs. Code generated by is competitive with optimized codes for matrix multiply on 256 nodes of the Lassen supercomputer and outperforms existing systems by between 1.8x to 3.7x (with a 45.7x outlier) on higher order tensor operations.

Task Fusion in Distributed Runtimes PDF
Shiv Sundram, Wonchan Lee and Alex Aiken
In the Parallel Applications Workshop, Alternatives to MPI+X (PAW-ATM 2022)
Abstract: We present distributed task fusion, a runtime optimization for task-based runtimes operating on parallel and heterogeneous systems. Distributed task fusion dynamically performs an efficient buffering, analysis, and fusion of asynchronously-evaluated distributed operations, reducing the overheads inherent to scheduling distributed tasks in implicitly parallel frameworks and runtimes. We identify the constraints under which distributed task fusion is permissible and describe an implementation in Legate, a domain-agnostic library for constructing portable and scalable task-based libraries. We present performance results using cuNumeric, a Legate library that enables scalable execution of NumPy pipelines on parallel and heterogeneous systems. We realize speedups up to 1.5x with task fusion enabled on up to 32 P100 GPUs, thus demonstrating efficient execution of pipelines involving many successive finegrained tasks. Finally, we discuss potential future work, including complementary optimizations that could result in additional performance improvements.

SpDISTAL: Compiling Distributed Sparse Tensor Computations PDF
Rohan Yadav, Alex Aiken and Fredrik Kjolstad
In the International Conference for High Performance Computing, Networking, Storage and Analysis (SC 2022)
Abstract: We introduce SpDISTAL, a compiler for sparse tensor algebra that targets distributed systems. SpDISTAL combines separate descriptions of tensor algebra expressions, sparse data structures, data distribution, and computation distribution. Thus, it enables distributed execution of sparse tensor algebra expressions with a wide variety of sparse data structures and data distributions. SpDISTAL is implemented as a C++ library that targets a distributed task-based runtime system and can generate code for nodes with both multi-core CPUs and multiple GPUs. SpDISTAL generates distributed code that achieves performance competitive with hand-written distributed functions for specific sparse tensor algebra expressions and that outperforms general interpretation-based systems by one to two orders of magnitude.

Visibility Algorithms for Dynamic Dependence Analysis and Distributed Coherence PDF
Michael Bauer, Elliott Slaughter, Sean Treichler, Wonchan Lee, Michael Garland and Alex Aiken
In Principles and Practices of Parallel Programming (PPoPP 2023)
Abstract: Implicitly parallel programming systems must solve the joint problems of dependence analysis and coherence to ensure apparently-sequential semantics for applications run on distributed memory machines. Solving these problems in the presence of data-dependent control flow and arbitrary aliasing is a challenge that most existing systems eschew by compromising the expressivity of their programming models and/or the performance of their implementations. We demonstrate a general class of solutions to these problems via a reduction to the visibility problem from computer graphics.

Legate Sparse: Distributed Sparse Computing in Python PDF
Rohan Yadav, Wonchan Lee, Melih Elibol, Taylor Lee Patti, Manolis Papadakis, Michael Garland, Alex Aiken, Fredrik Kjolstad and Michael Bauer
In the International Conference for High Performance Computing, Networking, Storage and Analysis (SC 2023)
Abstract: The sparse module of the popular SciPy Python library is widely used across applications in scientific computing, data analysis and machine learning. The standard implementation of SciPy is restricted to a single CPU and cannot take advantage of modern distributed and accelerated computing resources. We introduce Legate Sparse, a system that transparently distributes and accelerates unmodified sparse matrix-based SciPy programs across clusters of CPUs and GPUs, and composes with cuNumeric, a distributed NumPy library. Legate Sparse uses a combination of static and dynamic techniques to efficiently compose independently written sparse and dense array programming libraries, providing a unified Python interface for distributed sparse and dense array computations. We show that Legate Sparse is competitive with single-GPU libraries like CuPy and achieves 65% of the performance of PETSc on up to 1280 CPU cores and 192 GPUs of the Summit supercomputer, while offering the productivity benefits of idiomatic SciPy and NumPy.

Automated Mapping of Task-Based Programs onto Distributed and Heterogeneous Machines PDF
Thiago S. F. X. Teixeira, Alexandra Henzinger, Rohan Yadav and Alex Aiken
In the International Conference for High Performance Computing, Networking, Storage and Analysis (SC 2023)
Abstract: In a parallel and distributed application, a mapping is a selection of a processor for each computation or task and memories for the data collections that each task accesses. Finding high-performance mappings is challenging, particularly on heterogeneous hardware with multiple choices for processors and memories. We show that fast mappings are sensitive to the machine, application, and input. Porting to a new machine, modifying the application, or using a different input size may necessitate re-tuning the mapping to maintain the best possible performance.

We present AutoMap, a system that automatically tunes the mapping to the hardware used and finds fast mappings without user intervention or code modification. In contrast, hand-written mappings often require days of experimentation. AutoMap utilizes a novel constrained coordinate-wise descent search algorithm that balances the trade-off between running computations quickly and minimizing data movement. AutoMap discovers mappings up to 2.41× faster than custom, hand-written mappers.

Speaking Pygion: Experiences Writing an Exascale Single Particle Imaging Code PDF
Seema Mirchandaney, Alex Aiken, and Elliott Slaughter
In the Workshop on Asynchronous Many-Task Systems and Applications (WAMTA 2024)
Abstract: The goal of the SpiniFEL project was to write, from scratch, a single particle imaging code for exascale supercomputers. The original vision was to have two versions of the code, one in MPI and one in Pygion, a Python-based interface to the Legion task-based runtime. We describe the motivation for the project, some of the programming challenges we encountered along the way, what worked and what didn’t, and why only the Pygion code eventually succeeded in running at scale.

Theses

Note: The following thesis is a thorough guide to the Legion programming model and covers many implementation details that are not documented elsewhere.
Legion: Programming Distributed Heterogeneous Architectures with Logical Regions PDF
Michael Edward Bauer
December 2014
Abstract: This thesis covers the design and implementation of Legion, a new programming model and runtime system for targeting distributed heterogeneous machine architectures. Legion introduces logical regions as a new abstraction for describing the structure and usage of program data. We describe how logical regions provide a mechanism for applications to express important properties of program data, such as locality and independence, that are often ignored by current programming systems. We also show how logical regions allow programmers to scope the usage of program data by different computations. The explicit nature of logical regions makes these properties of programs manifest, allowing many of the challenging burdens of parallel programming, including dependence analysis and data movement, to be off-loaded from the programmer to the programming system.

Logical regions also improve the programmability and portability of applications by decoupling the specification of a program from how it is mapped onto a target architecture. Logical regions abstractly describe sets of program data without requiring any specification regarding the placement or layout of data. To control decisions about the placement of computations and data, we introduce a novel mapping interface that gives an application programmatic control over mapping decisions at runtime. Different implementations of the mapper interface can be used to port applications to new architectures and to explore alternative mapping choices. Legion guarantees that the decisions made through the mapping interface are independent of the correctness of the program, thus facilitating easy porting and tuning of applications to new architectures with different performance characteristics.

Using the information provided by logical regions, an implementation of Legion can automatically extract parallelism, manage data movement, and infer synchronization. We describe the algorithms and data structures necessary for efficiently performing these operations. We further show how the Legion runtime can be generalized to operate as a distributed system, making it possible for Legion applications to scale well. As both applications and machines continue to become more complex, the ability of Legion to relieve application developers of many of the tedious responsibilities they currently face will become increasingly important.

To demonstrate the performance of Legion, we port a production combustion simulation, called S3D, to Legion. We describe how S3D is implemented within the Legion programming model as well as the different mapping strategies that are employed to tune S3D for runs on different architectures. Our performance results show that a version of S3D running on Legion is nearly three times as fast as comparable state-of-the-art versions of S3D when run at 8192 nodes on the number two supercomputer in the world.

Realm: Performance Portability through Composable Asynchrony PDF
Sean Jeffrey Treichler
December 2016
Abstract: Modern supercomputers are growing increasingly complicated. The laws of physics have forced processor counts into the thousands or even millions, resulted in the creation of deep distributed memory hierarchies, and encouraged the use of multiple processor and memory types in the same system. Developing an application that is able to fully utilize such a system is very difficult. The development of an application that is able to run well on more than one such system with current programming models is so daunting that it is generally not even attempted.

The Legion project attempts to address these challenges by combining a traditional hierarchical application structure (i.e. tasks/functions calling other tasks/functions) with a hierarchical data model (logical regions, which may be partitioned into subregions), and introducing the concept of mapping, a process in which the tasks and regions of a machine-agnostic description are assigned to the processors and memories of a particular machine.

This dissertation focuses on Realm, the “low-level” runtime that manages the execution of a mapped Legion application. Realm is a fully asynchronous event-based runtime. Realm operations are deferred by the runtime, returning an event that triggers upon completion of the operation. These events may be used as preconditions for other operations, allowing arbitrary composition of asynchronous operations. The resulting operation graph naturally exposes the available parallelism in the application as well as opportunities for hiding the latency of any required communication. While asynchronous task launches and non-blocking data movement are fairly common in existing programming models, Realm makes all runtime operations asynchronous — this includes resource management, performance feedback, and even, apparently paradoxically, synchronization primitives.

Important design and implementation issues of Realm will be discussed, including the novel generational event data structure that allows Realm to efficiently and scalably handle a very large number of events in a distributed environment and the machine model that provides the information required for the mapping of a Legion application onto a system. Realm anticipates dynamic behavior of both future applications and future systems and includes mechanisms for application-directed profiling, fault reporting, and dynamic code generation that further improve performance portability by allowing an application to adapt to and optimize for the exact system configuration used for each run.

Microbenchmarks demonstrate the efficiency and scalability of the Realm and justify some of the non-obvious design decisions (e.g. unfairness in locks). Experiments with several mini-apps are used to measure the benefit of a fully asynchronous runtime compared to existing “non-blocking” approaches. Finally, performance of Legion applications at full-scale show how Realm’s composable asynchrony and support for heterogeneity benefit the overall Legion system on a variety of modern supercomputers.

Regent: A High-Productivity Programming Language for Implicit Parallelism with Logical Regions PDF
Elliott Slaughter
August 2017
Abstract: Modern supercomputers are dominated by distributed-memory machines. State of the art high-performance scientific applications targeting these machines are typically written in low-level, explicitly parallel programming models that enable maximal performance but expose the user to programming hazards such as data races and deadlocks. Conversely, implicitly parallel models isolate the user from these hazards by providing easy-to-use sequential semantics and place responsibility for parallelism and data movement on the system. However, traditional implementations of implicit parallelism suffer from substantial limitations: static, compiler-based implementations restrict the programming model to exclude dynamic features needed for unstructured applications, while dynamic, runtime-based approaches suffer from a sequential bottleneck that limits the scalability of the system.

We present Regent, a programming language designed to enable a hybrid static and dynamic analysis of implicit parallelism. Regent programs are composed of tasks (functions with annotated data usage). Program data is stored in regions (hierarchical collections); regions are dynamic, first-class values, but are named statically in the type system to ensure correct usage and analyzability of programs. Tasks may execute in parallel when they are mutually independent as determined by the annotated usage (read, write, etc.) of regions passed as task arguments. A Regent implementation is responsible for automatically discovering parallelism in a Regent program by analyzing the executed tasks in program order.

A naive implementation of Regent would suffer from a sequential bottleneck as tasks must be analyzed sequentially at runtime to discover parallelism, limiting scalability. We present an optimizing compiler for Regent which transforms implicitly parallel programs into efficient explicitly parallel code. By analyzing the region arguments to tasks, the compiler is able to determine the data movement implied by the sequence of task calls, even in the presence of unstructured and data-dependent application behavior. The compiler can then replace the implied data movement with explicit communication and synchronization for efficient execution on distributed-memory machines. We measure the performance and scalability of several Regent programs on large supercomputers and demonstrate that optimized Regent programs perform comparably to manually optimized explicitly parallel programs.

A Hybrid Approach to Automatic Program Parallelization via Efficient Tasking with Composable Data Partitioning PDF
Wonchan Lee
December 2019
Abstract: Despite the decades of research, distributed programming is still a painful task and programming systems designed to improve productivity fall short in practice. Auto-parallelizing compilers simplify distributed programming by parallelizing sequential programs automatically for distributed execution. However, their applicability is severely limited due to the fundamental undecidability of their static analysis problem. Runtime systems for implicit parallelism can handle a broader class of programs via an expressive programming model, but their runtime overhead often becomes a performance bottleneck. To design a practical system for productive distributed programming, one must combine the strengths of different parallelization paradigms to overcome their weaknesses when used in isolation.

This dissertation presents a hybrid approach to automatic program parallelization, which combines an auto-parallelizing compiler with an implicitly parallel tasking system. Our approach parallelizes programs in two steps. First, the auto-parallelizer materializes data parallelism in a program into task parallelism. Next, the tasking system dynamically analyzes dependencies between tasks and executes independent tasks in parallel. This two-stage process gives programmers a second chance when the auto-parallelizer “fails”: When a part of a program is not amenable to the compiler auto-parallelization, the programmer can gracefully fall back to the runtime parallelization by writing that part directly with task parallelism. Furthermore, hand-written tasks can be seamlessly integrated with the auto-parallelized part via composable data partitioning enabled by our auto-parallelizer, which allows them to share the partitioning strategy and thereby avoid excessive communication.

Key to the success of this hybrid approach is to minimize the overhead of the tasking system. To achieve this goal, we introduce dynamic tracing, a runtime mechanism for efficient tasking. The most expensive component in the tasking system is dynamic dependence analysis. Although this dynamic analysis is necessary when applications exhibit true dynamic behavior, the analysis is redundant for common cases where dependencies are (mostly) unchanging. Dynamic tracing eliminates this redundancy in dynamic dependence analysis by recording the dependence analysis of an execution trace and then replaying the recording for the subsequent occurrences of the same trace. To guarantee that a recording of a trace correctly replaces the trace’s original analysis, dynamic tracing also records memory locations that hold valid data when it records a trace and replays the recording only when those locations are still valid. Dynamic tracing significantly improves the efficiency of tasking, and thereby brings the strong scalability of explicit parallelism to implicit task parallelism.

Automated Discovery of Machine Learning Optimizations PDF
Zhihao Jia
August 2020
Abstract: The increasing complexity of machine learning (ML) models and ML-specific hardware architectures makes it increasingly challenging to build efficient and scalable ML systems. Today’s ML systems heavily rely on human effort to optimize the deployment of ML models on modern hardware platforms, which requires a tremendous amount of engineering effort but only provides suboptimal runtime performance. Moreover, the rapid evolution of ML models and ML-specific hardware makes it infeasible to manually optimize performance for all model and hardware combinations.

In this dissertation, we propose a search-based methodology to build performant ML systems by automatically discovering performance optimizations for ML computations. Instead of only considering the limited set of manually designed performance optimizations in current ML systems, our approach introduces a significantly more comprehensive search space of possible strategies to optimize the deployment of an ML model on a hardware platform. In addition, we design efficient search algorithms to explore the search space and discover highly-optimized strategies. The search is guided by a cost model for evaluating the performance of different strategies. We also propose a number of techniques to accelerate the search procedure by leveraging the topology of the search space.

This dissertation presents three ML systems that apply this methodology to optimize different tasks in ML deployment. Compared to current ML systems relying on manually designed optimizations, our ML systems enable better runtime performance by automatically discovering novel performance optimizations that are missing in current ML systems. Moreover, the performance improvement is achieved with less engineering effort, since the code needed for discovering these optimizations is much less than manual implementation of these optimizations.

First, we developed TASO, the first ML graph optimizer that automatically generates graph optimizations. TASO formally verifies the correctness of the generated graph optimizations using an automated theorem prover, and uses cost-based backtracking search to discover how to apply the verified optimizations. In addition to improving runtime performance and reducing engineering effort, TASO also provides correctness guarantees using formal methods.

Second, to generalize and go beyond today’s manually designed parallelization strategies for distributed ML computations, we introduce the SOAP search space, which contains a comprehensive set of possible strategies to parallelize ML computations by identifying parallelization opportunities across different Samples, Operators, Attributes, and Parameters. We developed FlexFlow, a deep learning engine that automatically searches over strategies in the SOAP search space. FlexFlow includes a novel execution simulator to evaluate the runtime performance of different strategies, and uses a Markov Chain Monte Carlo (MCMC) search algorithm to find performant strategies. FlexFlow discovers strategies that significantly outperform existing strategies, while requiring no manual effort during the search procedure.

Finally, we developed Roc, which automates data placement optimizations and minimizes data transfers in the memory hierarchy for large-scale graph neural network (GNN) computations. Roc formulates the task of optimizing data placement as a cost minimization problem and uses a dynamic programming algorithm to discover a globally optimal data management plan that minimizes data transfers between memories.

Scaling Implicit Parallelism with Index Launches PDF
Rupanshu Soi
December 2021
Abstract: Task-based programming systems are now widely used to program modern supercomputers. Since these systems need to achieve efficient scalable execution while being suitable for a wide variety of application domains, their core abstraction, the task graph, needs to have a scalable but expressive representation.

We present index launches, a technique, embedded in a general task-based framework, to collapse a task graph to obtain a more efficient representation. Leveraging a hybrid program analysis, we show how a compiler can support index launches by providing safety guarantees in virtually every case of practical importance.

In addition to the support for forall-style parallelism, we extend index launches in two primary directions. First, we introduce cross products, a novel abstraction for nested data partitioning, and describe how to incorporate them in index launches. Second, to enable index launches to represent limited forms of pipeline parallelism, we present ordered launches, and develop the program analysis required for an efficient implementation.

Our implementation of index launches is in Regent, a high-productivity programming language for distributed computing. Using the Piz Daint supercomputer, we evaluate the performance of several scientific applications written in Regent, and show that index launches lead to improved performance at the scales of contemporary and upcoming high-performance machines.