The Intel labs Haskell research compiler

The Glasgow Haskell Compiler (GHC) has quite sophisticated support for concurrency in its runtime system, which is written in low-level C code. As GHC evolves, the runtime system becomes increasingly complex, error-prone, difficult to maintain and difficult to add new concurrency features.This paper presents an alternative approach to implement concurrency in GHC. Rather than hard-wiring all kinds of concurrency features, the runtime system is a thin substrate providing only a small set of concurrency primitives, and the remaining concurrency features are implemented in software libraries written in Haskell. This design improves the safety of concurrency support; it also provides more customizability of concurrency features, which can be developed as Haskell library packages and deployed modularly.

Expressing algorithms using immutable arrays greatly simplifies the challenges of automatic SIMD vectorization, since several important classes of dependency violations cannot occur. The Haskell programming language provides libraries for programming with immutable arrays, and compiler support for optimizing them to eliminate the overhead of intermediate temporary arrays. We describe an implementation of automatic SIMD vectorization in a Haskell compiler which gives substantial vector speedups for a range of programs written in a natural programming style. We compare performance with that of programs compiled by the Glasgow Haskell Compiler.

A limited form of dependent types, called Generalized Algebraic Data Types (GADTs), has recently been added to the list of Haskell extensions supported by the Glasgow Haskell Compiler. Despite not being full-fledged dependent types, GADTs still offer considerably enlarged scope for enforcing important code and data invariants statically. Moreover, GADTs offer the tantalizing possibility of writing more efficient programs since capturing invariants statically through the type system sometimes obviates entire layers of dynamic tests and associated data markup. This paper is a case study on the applications of GADTs in the context of Yampa, a domain-specific language for Functional Reactive Programming in the form of a self-optimizing, arrow-based Haskell combinator library. The paper has two aims. Firstly, to explore what kind of optimizations GADTs make possible in this context. Much of that should also be relevant for other domain-specific embedded language implementations, in particular arrow-based ones. Secondly, as the actual performance impact of the GADT-based optimizations is not obvious, to quantify this impact, both on tailored micro benchmarks, to establish the effectiveness of individual optimizations, and on two fairly large, realistic applications, to gauge the overall impact. The performance gains for the micro benchmarks are substantial. This implies that the Yampa API could be simplified as a number of "pre-composed" primitives that were there mainly for performance reasons are no longer needed. As to the applications, a worthwhile performance gain was obtained in one case whereas the performance was more or less unchanged in the other.

We describe a monadic interface to low-level hardware features that is a suitable basis for building operating systems in Haskell. The interface includes primitives for controlling memory management hardware, user-mode process execution, and low-level device I/O. The interface enforces memory safety in nearly all circumstances. Its behavior is specified in part by formal assertions written in a programming logic called P-Logic. The interface has been implemented on bare IA32 hardware using the Glasgow Haskell Compiler (GHC) runtime system. We show how a variety of simple O/S kernels can be constructed on top of the interface, including a simple separation kernel and a demonstration system in which the kernel, window system, and all device drivers are written in Haskell.

Arrows are a popular form of abstract computation. Being more general than monads, they are more broadly applicable, and, in particular, are a good abstraction for signal processing and dataflow computations. Most notably, arrows form the basis for a domain-specific language called Yampa , which has been used in a variety of concrete applications, including animation, robotics, sound synthesis, control systems, and graphical user interfaces. Our primary interest is in better understanding the class of abstract computations captured by Yampa. Unfortunately, arrows are not concrete enough to do this with precision. To remedy this situation, we introduce the concept of commutative arrows that capture a noninterference property of concurrent computations. We also add an init operator that captures the causal nature of arrow effects, and identify its associated law. To study this class of computations in more detail, we define an extension to arrows called causal commutative arrows (CCA), and study its properties. Our key contribution is the identification of a normal form for CCA called causal commutative normal form (CCNF). By defining a normalization procedure, we have developed an optimization strategy that yields dramatic improvements in performance over conventional implementations of arrows. We have implemented this technique in Haskell, and conducted benchmarks that validate the effectiveness of our approach. When compiled with the Glasgow Haskell Compiler (GHC), the overall methodology can result in significant speedups.

The most widely available high performance platforms today are multilevel clusters of multicores. The Glasgow Haskell Compiler (GHC) provides a number of parallel Haskell implementations targeting different parallel architectures. In particular, GHC-SMP supports shared memory, and GHC-GUM supports distributed memory machines. Both implementations use different, but related, runtime environment (RTE) mechanisms. Good performance results can be achieved on shared memory architectures and on networks individually. However, a combination of both, for networks of multicores, is lacking.We present the design and implementation of a new parallel Haskell RTE implementation, GUMSMP, which exploits hierarchical platforms more effectively. It is designed to efficiently combine distributed memory parallelism, using a virtual shared heap over a cluster, with low-overhead shared memory parallelism on the multicores. Key design objectives for realising this system are: asymmetric load balance, effective latency hiding, and mostly passive load distribution.We show that the automatic hierarchical load distribution policies must be carefully tuned to obtain good performance, showing the impact of several policies, including work pre-fetching and favouring inter-node work distribution. We present the initial performance results for this implementation, demonstrating the good scalability of a set of 8 benchmarks on up to 100 cores, and show performance gains of up to 20% compared to GHC-GUM.

We have developed a new, portable I/O event manager for the Glasgow Haskell Compiler (GHC) that scales to the needs of modern server applications. Our new code is transparently available to existing Haskell applications. Performance at lower concurrency levels is comparable with the existing implementation. We support millions of concurrent network connections, with millions of active timeouts, from a single multithreaded program, levels far beyond those achievable with the current I/O manager. In addition, we provide a public API to developers who need to create event-driven network applications.

The problems of stage-preserving linearization and one-boundedness are studied for a class of nonlinear single rule recursive programs, and syntactic characterizations are developed for both. The characterizations lead to a polynomial-time algorithm for the former and a linear-time algorithm for the latter. Stage-preserving linearization results in a significant improvement in evaluation efficiency, compared to a linearization that does not preserve stages. The class of nonlinear strips that are stage-preserving linearizable includes several classes of programs that can be linearized only using a mix of left and right linear rules, as well as programs that cannot be linearized using previously known techniques. The study makes use of a technique based on the notion of homomorphic tree embeddings. >

In an impure functional language, there are programs whose behaviour is completely functional (in that they behave extensionally on inputs), but the functions they compute cannot be written in the purely functional fragment of the language. That is, the class of programs with functional behaviour is more expressive than the usual class of pure functional programs. In this paper we introduce this extended class of "functional" programs by means of examples in Standard ML, and explore what they might have to offer to programmers and language implementors.After reviewing some theoretical background, we present some examples of functions of the above kind, and discuss how they may be implemented. We then consider two possible programming applications for these functions: the implementation of a search algorithm, and an algorithm for exact real-number integration. We discuss the advantages and limitations of this style of programming relative to other approaches. We also consider the increased scope for compiler optimizations that these functions would offer.

This volume contains the proceedings of the Sixth Workshop on the Implementation, Compilation, Optimization of Object-Oriented Languages, Programs and Systems (ICOOOLPS). Computer programming languages, especially Object-Oriented languages, are pervasive and play a significant role in computer science and engineering life and sometime appear as ubiquitous and completely mature. However, despite a large number of works, there still is a clear need for solutions for efficient implementation and compilation of OO languages in various application domains ranging from embedded and real-time systems to desktop systems. The ICOOOLPS workshop thus aims to address this crucial issue of optimization, mainly but not only in OO languages, programs and systems. It intends to do so by bringing together researchers and practitioners working in the field of implementation and optimization of programming languages. Its main goals are identifying fundamental bases and key current issues pertaining to the efficient implementation, compilation and optimization of OO languages, and outlining future challenges and research directions.

The purpose of this thesis is to study the composition of datalog mappings in order to analyze the situation where a sequence of datalog program queries are evaluated serially. Four related topics are investigated and shown to be useful for the optimization of datalog programs. First, the problem of when a composition mapping is equivalent to a datalog mapping (composability) is considered. An algebraic structure called composition system is abstracted from datalog mappings, and used to establish three characterizations for composability. Several consequences of these characterizations, including invariance properties of composability, are presented. Second, the reverse process of composition, namely decomposition, is examined. A canonical decomposition into (usually) smaller programs is given, as well as the decomposition of strongly linear programs and bounded programs into single-rule programs. Third, the class of prime programs is introduced and then scrutinized. Major results include the primality of a class of single-rule programs called symmetric; the existence of arbitrarily large primes; the factorization of bounded programs into single-rule primes; and a condition for the uniqueness of factorizations into primes. And fourth, a structure called datalog module is investigated in order to understand the infinite set of compositions of mappings formed from a finite set of programs. The family of datalog modules is shown to be a composition system. Finally, a special class of datalog modules, called derivation-bounded, is introduced and proven to have efficient computation and storage features. (Copies available exclusively from Micrographics Department, Doheny Library, USC, University Park, Los Angeles, CA 90089-0182.)

A large part of the development effort of compute-intensive applications is devoted to optimization, i.e., achieving the computation within a finite budget of time, space or energy. Given the complexity of modern architectures, writing simulation applications is often a two-step workflow. Firstly, developers design a sequential program for algorithmic tuning and debugging purposes. Secondly, experts optimize and exploit possible approximations of the original program to scale to the actual problem size. This second step is a tedious, time-consuming and error-prone task. In this paper we investigate language extensions and compiler tools to achieve that task semi-automatically in the context of approximate computing. We identified the semantic and syntactic information necessary for a compiler to automatically handle approximation and adaptive techniques for a particular class of programs. We propose a set of language extensions generic enough to provide the compiler with the useful semantic information when approximation is beneficial. We implemented the compiler infrastructure to exploit these extensions and to automatically generate the adaptively approximated version of a program. We provide an experimental study of the impact and expressiveness of our language extension set on various applications.

Robust optimization and stochastic programming approaches for medium-term production scheduling of a large-scale steelmaking continuous casting process under demand uncertainty

In this paper, we study multistage stochastic mixed-integer nonlinear programs (MS-MINLP). This general class of problems encompasses, as important special cases, multistage stochastic convex optimization with non-Lipschitzian value functions and multistage stochastic mixed-integer linear optimization. We develop stochastic dual dynamic programming (SDDP) type algorithms with nested decomposition, deterministic sampling, and stochastic sampling. The key ingredient is a new type of cuts based on generalized conjugacy. Several interesting classes of MS-MINLP are identified, where the new algorithms are guaranteed to obtain the global optimum without the assumption of complete recourse. This significantly generalizes the classic SDDP algorithms. We also characterize the iteration complexity of the proposed algorithms. In particular, for a (T+1)-stage stochastic MINLP satisfying L-exact Lipschitz regularization with d-dimensional state spaces, to obtain an varepsilon -optimal root node solution, we prove that the number of iterations of the proposed deterministic sampling algorithm is upper bounded by {mathcal {O}}((frac{2LT}{varepsilon })^d), and is lower bounded by {mathcal {O}}((frac{LT}{4varepsilon })^d) for the general case or by {mathcal {O}}((frac{LT}{8varepsilon })^{d/2-1}) for the convex case. This shows that the obtained complexity bounds are rather sharp. It also reveals that the iteration complexity depends polynomially on the number of stages. We further show that the iteration complexity depends linearly on T, if all the state spaces are finite sets, or if we seek a (Tvarepsilon )-optimal solution when the state spaces are infinite sets, i.e. allowing the optimality gap to scale with T. To the best of our knowledge, this is the first work that reports global optimization algorithms as well as iteration complexity results for solving such a large class of multistage stochastic programs. The iteration complexity study resolves a conjecture by the late Prof. Shabbir Ahmed in the general setting of multistage stochastic mixed-integer optimization.

The complexity-performance trade-off is a fundamental aspect of the design of low-density parity-check (LDPC) codes. In this paper, we consider LDPC codes for the binary erasure channel (BEC), use code rate for performance metric, and number of decoding iterations to achieve a certain residual erasure probability for complexity metric. We first propose a quite accurate approximation of the number of iterations for the BEC. Moreover, a simple but efficient utility function corresponding to the number of iterations is developed. Using the aforementioned approximation and the utility function, two optimization problems w.r.t. complexity are formulated to find the code degree distributions. We show that both optimization problems are convex. In particular, the problem with the proposed approximation belongs to the class of semi-infinite problems which are computationally challenging to be solved. However, the problem with the proposed utility function falls into the class of semi-definite programming (SDP) and thus, the global solution can be found efficiently using available SDP solvers. Numerical results reveal the superiority of the proposed code design compared to existing code designs from literature.