Modern programming languages offer special syntax and semantics for logical fork-join parallelism in the form of parallel loops, allowing them to be nested, e.g., a parallel loop within another parallel loop. This expressiveness comes at a price, however: on modern multicore systems, realizing logical parallelism results in overheads due to the creation and management of parallel tasks, which can wipe out the benefits of parallelism. Today, we expect application programmers to cope with it by manually tuning and optimizing their code. Such tuning requires programmers to reason about architectural factors hidden behind layers of software abstractions, such as task scheduling and load balancing. Managing these factors is particularly challenging when workloads are irregular because their performance is input-sensitive. This paper presents HBC, the first compiler that translates C/C++ programs with high-level, fork-join constructs (e.g., OpenMP) to binaries capable of automatically controlling the cost of parallelism and dealing with irregular, input-sensitive workloads. The basis of our approach is Heartbeat Scheduling, a recent proposal for automatic granularity control, which is backed by formal guarantees on performance. HBC binaries outperform OpenMP binaries for workloads for which even entirely manual solutions struggle to find the right balance between parallelism and its costs.

Over the years, traditional tracing garbage collectors have accumulated assumptions that may not hold in new language designs. For instance, we usually assume that run-time objects do not hold addressable sub-parts and have a size of at least one pointer. These fail in systems striving to eliminate pointers and represent data in a dense, serialized form, such as the Gibbon compiler. We propose a new memory management strategy for language runtimes with mostly serialized heaps. It uses a hybrid, generational collector, where regions are bump-allocated into the young generation and objects are bump-allocated within those regions. Minor collections copy data into larger regions in the old generation, compacting it further. The old generation uses region-level reference counting. The resulting system maintains high performance for data traversal programs, while significantly improving performance on other kinds of allocation patterns.

On any modern computer architecture today, parallelism comes with a modest cost, born from the creation and management of threads or tasks. Today, programmers battle this cost by manually optimizing/tuning their codes to minimize the cost of parallelism without harming its benefit, performance. This is a difficult battle: programmers must reason about architectural constant factors hidden behind layers of software abstractions, including thread schedulers and memory managers, and their impact on performance, also at scale. In languages that support higher-order functions, the battle hardens: higher order functions can make it difficult, if not impossible, to reason about the cost and benefits of parallelism. Motivated by these challenges and the numerous advantages of high-level languages, we believe that it has become essential to manage parallelism automatically so as to minimize its cost and maximize its benefit. This is a challenging problem, even when considered on a case-by-case, application-specific basis. But if a solution were possible, then it could combine the many correctness benefits of high-level languages with performance by managing parallelism without the programmer effort needed to ensure performance. This paper proposes techniques for such automatic management of parallelism by combining static (compilation) and run-time techniques. Specifically, we consider the Parallel ML language with task parallelism, and describe a compiler pipeline that embeds “potential parallelism” directly into the call-stack and avoids the cost of task creation by default. We then pair this compilation pipeline with a run-time system that dynamically converts potential parallelism into actual parallel tasks. Together, the compiler and run-time system guarantee that the cost of parallelism remains low without losing its benefit. We prove that our techniques have no asymptotic impact on the work and span of parallel programs and thus preserve their asymptotic properties. We implement the proposed techniques by extending the MPL compiler for Parallel ML and show that it can eliminate the burden of manual optimization while delivering good practical performance.

Programming languages using functions on collections of values, such as map, reduce, scan and filter, have been used for over fifty years. Such collections have proven to be particularly useful in the context of parallelism because such functions are naturally parallel. However, if implemented naively they lead to the generation of temporary intermediate collections that can significantly increase memory usage and runtime. To avoid this pitfall, many approaches use "fusion" to combine operations and avoid temporary results. However, most of these approaches involve significant changes to a compiler and are limited to a small set of functions, such as maps and reduces.

Achieving parallel performance and scalability involves making compromises between parallel and sequential computation. If not contained, the overheads of parallelism can easily outweigh its benefits, sometimes by orders of magnitude. Today, we expect programmers to implement this compromise by optimizing their code manually. This process is labor intensive, requires deep expertise, and reduces code quality. Recent work on heartbeat scheduling shows a promising approach that manifests the potentially vast amounts of available, latent parallelism, at a regular rate, based on even beats in time. The idea is to amortize the overheads of parallelism over the useful work performed between the beats. Heartbeat scheduling is promising in theory, but the reality is complicated: it has no known practical implementation.

Recent work showed that compiling functional programs to use dense, serialized memory representations for recursive algebraic datatypes can yield significant constant-factor speedups for sequential programs. But serializing data in a maximally dense format consequently serializes the processing of that data, yielding a tension between density and parallelism. This paper shows that a disciplined, practical compromise is possible. We present Parallel Gibbon, a compiler that obtains the benefits of dense data formats and parallelism. We formalize the semantics of the parallel location calculus underpinning this novel implementation strategy, and show that it is type-safe. Parallel Gibbon exceeds the parallel performance of existing compilers for purely functional programs that use recursive algebraic datatypes, including, notably, abstract-syntax-tree traversals as in compilers.

In a typical data-processing program, the representation of data in memory is distinct from its representation in a serialized form on disk. The former has pointers and arbitrary, sparse layout, facilitating easy manipulation by a program, while the latter is packed contiguously, facilitating easy I/O. We propose a language, LoCal, to unify in-memory and serialized formats. LoCal extends a region calculus into a location calculus, employing a type system that tracks the byte-addressed layout of all heap values. We formalize LoCal and prove type safety, and show how LoCal programs can be inferred from unannotated source terms.

Over the past decade, many programming languages and systems for parallel-computing have been developed, e.g., Fork/Join and Habanero Java, Parallel Haskell, Parallel ML, and X10. Although these systems raise the level of abstraction for writing parallel codes, performance continues to require labor-intensive optimizations for coarsening the granularity of parallel executions. In this paper, we present provably and practically efficient techniques for controlling granularity within the run-time system of the language. Our starting point is "oracle-guided scheduling", a result from the functional-programming community that shows that granularity can be controlled by an "oracle" that can predict the execution time of parallel codes. We give an algorithm for implementing such an oracle and prove that it has the desired theoretical properties under the nested-parallel programming model. We implement the oracle in C++ by extending Cilk and evaluate its practical performance. The results show that our techniques can essentially eliminate hand tuning while closely matching the performance of hand tuned codes.

A classic problem in parallel computing is to take a high-level parallel program written, for example, in nested-parallel style with fork-join constructs and run it efficiently on a real machine. The problem could be considered solved in theory, but not in practice, because the overheads of creating and managing parallel threads can overwhelm their benefits. Developing efficient parallel codes therefore usually requires extensive tuning and optimizations to reduce parallelism just to a point where the overheads become acceptable.

Over the past decade, many programming languages and systems for parallel-computing have been developed, including Cilk, Fork/Join Java, Habanero Java, Parallel Haskell, Parallel ML, and X10. Although these systems raise the level of abstraction at which parallel code are written, performance continues to require the programmer to perform extensive optimizations and tuning, often by taking various architectural details into account. One such key optimization is granularity control, which requires the programmer to determine when and how parallel tasks should be sequentialized.

Over the past two decades, many concurrent data structures have been designed and implemented. Nearly all such work analyzes concurrent data structures empirically, omitting asymptotic bounds on their efficiency, partly because of the complexity of the analysis needed, and partly because of the difficulty of obtaining relevant asymptotic bounds: when the analysis takes into account important practical factors, such as contention, it is difficult or even impossible to prove desirable bounds.

Increasing availability of multicore systems has led to greater focus on the design and implementation of languages for writing parallel programs. Such languages support various abstractions for parallelism, such as fork-join, async-finish, futures. While they may seem similar, these abstractions lead to different semantics, language design and implementation decisions, and can significantly impact the performance of end-user applications.

Abstract A classic problem in parallel computing is determining whether to execute a thread in parallel or sequentially. If small threads are executed in parallel, the overheads due to thread creation can overwhelm the benefits of parallelism, resulting in suboptimal efficiency and performance. If large threads are executed sequentially, processors may spin idle, resulting again in sub-optimal efficiency and performance. This “granularity problem” is especially important in implicitly parallel languages, where the programmer expresses all potential for parallelism, leaving it to the system to exploit parallelism by creating threads as necessary. Although this problem has been identified as an important problem, it is not well understood—broadly applicable solutions remain elusive. In this paper, we propose techniques for automatically controlling granularity in implicitly parallel programming languages to achieve parallel efficiency and performance. To this end, we first extend a classic result, Brent's theorem (a.k.a. the work-time principle) to include thread-creation overheads. Using a cost semantics for a general-purpose language in the style of lambda calculus with parallel tuples, we then present a precise accounting of thread-creation overheads and bound their impact on efficiency and performance. To reduce such overheads, we propose an oracle-guided semantics by using estimates of the sizes of parallel threads. We show that, if the oracle provides accurate estimates in constant time, then the oracle-guided semantics reduces the thread-creation overheads for a reasonably large class of parallel computations. We describe how to approximate the oracle-guided semantics in practice by combining static and dynamic techniques. We require the programmer to provide the asymptotic complexity cost for each parallel thread and use runtime profiling to determine hardware-specific constant factors. We present an implementation of the proposed approach as an extension of the Manticore compiler for Parallel ML. Our empirical evaluation shows that our techniques can reduce thread-creation overheads, leading to good efficiency and performance.

Advances in processing power and memory technology have made multicore computers an important platform for high-performance graph-search (or graph-traversal) algorithms. Since the introduction of multicore, much progress has been made to improve parallel breadth-first search. However, less attention has been given to algorithms for unordered or loosely ordered traversals.

Work stealing has proven to be an effective method for scheduling parallel programs on multicore computers. To achieve high performance, work stealing distributes tasks between concurrent queues, called deques, which are assigned to each processor. Each processor operates on its deque locally except when performing load balancing via steals. Unfortunately, concurrent deques suffer from two limitations: 1) local deque operations require expensive memory fences in modern weak-memory architectures, 2) they can be very difficult to extend to support various optimizations and flexible forms of task distribution strategies needed many applications, e.g., those that do not fit nicely into the divide-and-conquer, nested data parallel paradigm.

Data parallelism has proven to be an effective technique for high-level programming of a certain class of parallel applications, but it is not well suited to irregular parallel computations. Blelloch and others proposed nested data parallelism (NDP) as a language mechanism for programming irregular parallel applications in a declarative data-parallel style. The key to this approach is a compiler transformation that flattens the NDP computation and data structures into a form that can be executed efficiently on a wide-vector SIMD architecture. Unfortunately, this technique is ill suited to execution on today's multicore machines. We present a new technique, called data-only flattening, for the compilation of NDP, which is suitable for multicore architectures. Data-only flattening transforms nested data structures in order to expose programs to various optimizations while leaving control structures intact. We present a formal semantics of data-only flattening in a core language with a rewriting system. We demonstrate the effectiveness of this technique in the Parallel ML implementation and we report encouraging experimental results across various benchmark applications.

Abstract Nested data-parallelism (NDP) is a language mechanism that supports programming irregular parallel applications in a declarative style. In this paper, we describe the implementation of NDP in Parallel ML (PML), which is a part of the Manticore system. One of the main challenges of implementing NDP is managing the parallel decomposition of work. If we have too many small chunks of work, the overhead will be too high, but if we do not have enough chunks of work, processors will be idle. Recently, the technique of Lazy Binary Splitting was proposed to address this problem for nested parallel loops over flat arrays. We have adapted this technique to our implementation of NDP, which uses binary trees to represent parallel arrays. This new technique, which we call Lazy Tree Splitting (LTS), has the key advantage of performance robustness , i.e., it does not require tuning to get the best performance for each program. We describe the implementation of the standard NDP operations using LTS and present experimental data that demonstrate the scalability of LTS across a range of benchmarks.

A classic problem in parallel computing is determining whether to execute a task in parallel or sequentially. If small tasks are executed in parallel, the task-creation overheads can be overwhelming. If large tasks are executed sequentially, processors may spin idle. This granularity problem, however well known, is not well understood: broadly applicable solutions remain elusive.

Nested data-parallelism (NDP) is a declarative style for programming irregular parallel applications. NDP languages provide language features favoring the NDP style, efficient compilation of NDP programs, and various common NDP operations like parallel maps, filters, and sum-like reductions. In this paper, we describe the implementation of NDP in Parallel ML (PML), part of the Manticore project. Managing the parallel decomposition of work is one of the main challenges of implementing NDP. If the decomposition creates too many small chunks of work, performance will be eroded by too much parallel overhead. If, on the other hand, there are too few large chunks of work, there will be too much sequential processing and processors will sit idle.

Abstract The increasing availability of commodity multicore processors is making parallel computing ever more widespread. In order to exploit its potential, programmers need languages that make the benefits of parallelism accessible and understandable. Previous parallel languages have traditionally been intended for large-scale scientific computing, and they tend not to be well suited to programming the applications one typically finds on a desktop system. Thus, we need new parallel-language designs that address a broader spectrum of applications. The Manticore project is our effort to address this need. At its core is Parallel ML, a high-level functional language for programming parallel applications on commodity multicore hardware. Parallel ML provides a diverse collection of parallel constructs for different granularities of work. In this paper, we focus on the implicitly threaded parallel constructs of the language, which support fine-grained parallelism. We concentrate on those elements that distinguish our design from related ones, namely, a novel parallel binding form, a nondeterministic parallel case form, and the treatment of exceptions in the presence of data parallelism. These features differentiate the present work from related work on functional data-parallel language designs, which have focused largely on parallel problems with regular structure and the compiler transformations—most notably, flattening—that make such designs feasible. We present detailed examples utilizing various mechanisms of the language and give a formal description of our implementation.

The trend in microprocessor design toward multicore and manycore processors means that future performance gains in software will largely come from harnessing parallelism. To realize such gains, we need languages and implementations that can enable parallelism at many different levels. For example, an application might use both explicit threads to implement course-grain parallelism for independent tasks and implicit threads for fine-grain data-parallel computation over a large array. An important aspect of this requirement is supporting a wide range of different scheduling mechanisms for parallel computation.