Non-numeric Applications Research Articles

Understanding why correlation profiling improves the predictability of data cache misses in nonnumeric applications

Latency-tolerance techniques offer the potential for bridging the ever-increasing speed gap between the memory subsystem and today's high-performance processors. However, to fully exploit the benefit of these techniques, one must be careful to apply them only to the dynamic references that are likely to suffer cache misses-otherwise the runtime overheads can potentially offset any gains. In this paper, we focus on isolating dynamic miss instances in nonnumeric applications, which is a difficult but important problem. Although compilers cannot statically analyze data locality in nonnumeric applications, one viable approach is to use profiling information to measure the actual miss behavior. Unfortunately, the state-of-the-art in cache miss profiling (which we call summary profiling) is inadequate for references with intermediate miss ratios-it either misses opportunities to hide latency, or else inserts overhead that is unnecessary. To overcome this problem, we propose and evaluate a new profiling technique that helps predict which dynamic instances of a static memory reference will hit or miss in the cache: correlation profiling Our experimental results demonstrate that roughly half of the 21 nonnumeric applications we study can potentially enjoy significant reductions in memory stall time by exploiting at least one of the three forms of correlation profiling we consider: control-flow correlation, self correlation, and global correlation. In addition, our detailed case studies illustrate that self correlation succeeds because a given reference's cache outcomes often contain repeated patterns and control-flow correlation succeeds because cache outcomes are often call-chain dependent. Finally, we suggest a number of ways to exploit correlation profiling in practice and demonstrate that software prefetching can achieve better performance on a modern superscalar processor when directed by correlation profiling rather than summary profiling information.

Experiences with non-numeric applications on multithreaded architectures

Distributed-memory machines have proved successful for many challenging numerical programs that can be split into largely independent computation-intensive subtasks requiring little data exchange (although the amount of exchanged data may be large). However, many irregular applications---e.g. in the AI field --- have a fairly tight data coupling that often results from the use of shared data structures, making them in many cases not amenable to parallelization on distributed-memory machines. EARTH is an efficient multithreaded architecture that supports in particular large numbers of small data exchanges by means of low start-up times and the ability of tolerance of even small latencies. In this paper, we show the benefits provided by EARTH for applications of this sort by presenting experimental results from several AI applications run on the MANNA machine, which is a distributed-memory machine with a very high-performance communicantion network. EARTH-MANNA is shown to extend the range of programs that can be parallelized and run effectively on distributed-memory machines.

Available instruction-level parallelism for superscalar and superpipelined machines

Superscalar machines can issue several instructions per cycle. Superpipelined machines can issue only one instruction per cycle, but they have cycle times shorter than the latency of any functional unit. In this paper these two techniques are shown to be roughly equivalent ways of exploiting instruction-level parallelism. A parameterizable code reorganization and simulation system was developed and used to measure instruction-level parallelism for a series of benchmarks. Results of these simulations in the presence of various compiler optimizations are presented. The average degree of superpipelining metric is introduced. Our simulations suggest that this metric is already high for many machines. These machines already exploit all of the instruction-level parallelism available in many non-numeric applications, even without parallel instruction issue or higher degrees of pipelining.

An Algorithm to Decide Feasibility of Linear Integer Constraints Occurrng in Decision Tables

To detect errors in decision tables one needs to decide whether a given set of constraints is feasible or not. This paper describes an algorithm to do so when the constraints are linear in variables that take only integer values. Decision tables with such constraints occur frequently in business data processing and in nonnumeric applications. The aim of the algorithm is to exploit. the abundance of very simple constraints that occur in typical decision table contexts. Essentially, the algorithm is a backtrack procedure where the the solution space is pruned by using the set of simple constrains. After some simplications, the simple constraints are captured in an acyclic directed graph with weighted edges. Further, only those partial vectors are considered from extension which can be extended to assignments that will at least satisfy the simple constraints. This is how pruning of the solution space is achieved. For every partial assignment considered, the graph representation of the simple constraints provides a lower bound for each variable which is not yet assigned a value. These lower bounds play a vital role in the algorithm and they are obtained in an efficient manner by updating older lower bounds. Our present algorithm also incorporates an idea by which it can be checked whether or not an (m –2)-ary vector can be extended to a solution vector of m components, thereby backtracking is reduced by one component.

Nonnumeric Computer Applications to Algebra, Trigonometry and Calculus

Algebra. Early automatic simplification programs employed an assemblage of ad hoc techniques such as collecting similar terms or factors, and employing transformations such as for all u: 1 U U, Ou*0, 0+ U--u. This approach is adequate for elementary simplification. However, the more sophisticated operations such as cancellation of polynomial greatest common divisors, factorization, solution of nonlinear equations, and simplification of radicals are far more efficient and easier to implement using a more systematic algebraic approach. Diverse topics make important contributions here. Particularly relevant topics are ring theory (including commutative, noncommutative and principal ideals), integral domains, Euclidean domains, unique factorization domains, field theory (including Galois, finite and quotient fields), algebraic geometry (including singular perturbations and catastrophe theory), p-adic numbers and valuation theory. Conversely, computer algebra can serve as a powerful tool for abstract algebra: Some systems have facilities that make it particularly easy to implement new algebraic structures, thus facilitating an empirical (perhaps exhaustive) study of some of their properties. For example, MACSYMA permits one to establish a new operator and declare that it has certain properties such as commutativity, associativity, left distributivity, etc. MACSYMA also has a so-called pattern matcher that makes it easy to establish automatic substitution rules such as q -k and k2 2_ 1. With the help of these mechanisms it is particularly easy to implement an additional known algebraic structure (such as quaternions) or to implement an original algebraic structure in order to help study its properties. MACSYMA also has a supplementary package that provides particularly strong support for investigations of commutative rings. For example, to establish R as the ring of polynomials in x, y, z and w over the integers mod 231 1, mod Z2 _ W3, one merely enters the commands

Evaluation of alternate data base machine designs

The purpose of this paper is to point out the need for performance evaluation measures and techniques suitable for the evaluation of specialized architectural features in nonnumeric applications. Toward this end, problems associated with the use of data base machines are examined at three levels of detail: the user level, the system level and the device level.

The Architectural Features and Implementation Techniques of the Multicell CASSM

The architectural characteristics and the implementation techniques of a context addressed segment sequential memory system called CASSM are described. The system provides hardware support for many database management functions. It offers associative and parallel processing capabilities for the efficient retrieval and manipulation of data in large databases. The hardware is designed mainly to support a hierarchical model for database applications but also contains facilities for supporting a wide range of data searches and operations useful for other nonnumeric data processing applications. The software development of the assembly language CASAL and its assembler, the high-level nonprocedural language CASDAL and its compiler, the interface to the CASSM-user interface computer, etc., have been carried out for the system. The hardware design and implementation techniques have been verified using a one-cell prototype system and a simulator designed for testing the system in the multicell environment. The emphasis of this paper is in the detailed description of the hardware features and techniques used in the multicell CASSM system.

Data base machines

In the past few years there has been increased emphasis in the use of new and different hardware in the data base management field. This has given rise to an interesting area of research called data base machines. The objectives are to implement many of the traditional software data base management functions in hardware in order to increase performance and capability.The reasons for increased research activity in data base machines are many fold. The modern digital computer with all its power still executes only a single instruction at a time al beit with a few exceptions. Since the digital computer was primarily designed for numeric applications, the ratio of useful instructions to overhead instructions tends to be very low in non-numeric applications such as data base management. There is an increasing body of knowledge that indicates that parallelism and addressing by content can be used to advantage in such data base management functions as retrieval and update. The above reasons coupled with the fact that software and personnel costs continued to rise while hardware costs continue to decrease makes it imperative that researchers find new solutions to the data base management problem that take maximum advantage of hardware architecture and technology.In this paper a brief review of current efforts in data base machines is presented. A review of the more important functions performed in data base management is given first. This is followed by a brief discussion of mini computer implementations. Then a review of some of the current efforts in data base machines is given. Finally, a few thoughts are presented on future trends.

A comparison of Cobol, Fortran, PL-I and Spitbol

Spitbol, the compiler version of the symbol manipulation language Snobol 4, is compared with the well-known higher-level languages of Cobol, Fortran and PL-I. An algorithm for searching and insertion into a binary tree was chosen for a comparison standard and programmed in each language. Development and computer costs were then compared for the four languages. Although a single comparison does not prove one language superior to another, this comparison does suggest that Spitbol should at least be considered when choosing a programming language for non-numeric applications.

Implementation of a context-addressed pipeline 3 SIMD architecture

The context-addressed SIMD architecture is being examined as a means to provide suitable processing capabilities for non-numeric applications. This paper considers techniques which are being applied to realize this architecture in a segment sequential fashion.A fixed head disc provides a large amount of sequential storage already segmented into tracks. Economical bit-serial microprocessors are employed one per track to process the instructions and data sequentially as the disc rotates. A microprocessor and its track are here called a cell. A chain of cells can search a file consisting of fixed length words organized into variable length records in one revolution or cycle of the disc. Intercell communication in the chain as a pipeline allows files and records which overlap segment boundaries to be processed as an intact file or record, with all segments processed in parallel. Transfers of words and bits in pipeline fashion between cells allows the file to change size dynamically without regard for its cellular segmentation, as words may be either inserted or deleted within a record.Instruction execution in this architecture is also pipelined. The next instruction is preprocessed in one cycle along with its operand while the current instruction is executing, and the previous one is finishing up its operation or post-processing. This instruction pipelining is performed in order to minimize the execution time for a sequential instruction by overlapping the instruction fetch and execute of two consecutive cycles. The segment sequential organization also demands that the instruction execution continue into the next cycle in order to process the elements of a record that may overlap segments. Again by overlapping this instruction clean-up with the next instruction's execution, we may minimize the instruction execution time.A hardware realization of this pipeline of cells which pipelines its instruction execution becomes another pipeline. Each bit serial microprocessor is in fact a pipeline of modules each performing some phase of instruction execution or segment support such as garbage word deletion or new data insertion. The segment sequential architecture lends itself very nicely to this pipeline 3 approach.Although they provide a powerful architecture, these three pipeline constructs also develope new problems in their control and timing. Such problems include that the words which are pipelined between cells are not processed twice, missed entirely, or only partially processed by the instruction pipeline. A solution to these and to other problems related to this architecture are presented in this paper.

Nonnumeric applications of microprocessors

Two nonnumeric applications of microprocessors are described. It is shown that commercially available products allow an economical design of a small private automatic branch exchange (PABX). The second application considers a laboratory data aquisition system. Microprocessor characteristics are assessed on the basis of our experience with these applications. In view of some of the shortcomings of current products, some sugsestions are made which would make microprocessors more suitable for the nonnumeric environment. To demonstrate the feasibility of these suggestions, a design of a suitable micro-processor is proposed.

A code for non-numeric information processing applications in online systems

A code has been specifically designed to simplify the internal information processing operations within an online computer system with respect to non-numeric applications, and to maximize the transfer rate of the information channel linking the system and the system user. The code has direct application to problems in areas such as information retrieval, document classification, computer-aided teaching and text editing. This code, called IPC (Information Processing Code), is an 8-bit code set constructed so that 7, 6, 5 and 4-bit subsets can be easily derived from the basic set. The code set is organized so that simple binary operations can distinguish between the numeric alphabetic, special symbol and control character codes. The number of useable characters within the basic set size may be expanded either by use of escape codes included in the set, or by suitable interpretation of otherwise unassigned codes on the basis of the requirements of local environments.

The Digital Computer as an Aid in Linguistics

1. Introductory. The recent rapid development of electronic computing facilities on college and university campuses has become particularly dramatic during the past several months. Dozens of colleges and universities now have 'medium-sized' computers such as the IBM 650, and several campuses have larger machines. The best known of the large-scale digital computers are the IBM 704, 709, and 7090 systems, which are very similar to each other. Among the universities which have machines of this type are the Massachusetts Institute of Technology and the Universities of Michigan, Washington, and California (at both Berkeley and Los Angeles), and several additional ones have recently been installed or are on order at other institutions. Comparable machines made by Remington-Rand are available at Harvard University, the University of Pennsylvania, and the University of North Carolina. The installations on these campuses exist for research and educational purposes; in other words, they are intended to be of service to the faculty and students. This being the case, it is standard practice to make the machines available free of charge to faculty members and students who have no outside financial support for their research, provided their projects are deemed worth while. This fortunate circumstance is extremely significant, since the use of these machines is ordinarily very costly. For example, the standard rental for a typical 704 is about $45,000 per month, and the price for its use on an hourly basis, if it were a commercial installation, would be around $300 to $400 per hour, which amounts to about $6 per minute or ten cents per second. Also important is the fact that the universities which have these large computers generally provide free use of them to faculty members of other academic institutions in their geographic areas, as long as time is available. This means that most linguists in the United States are potentially in a position to make use of large-scale computers. The reason that computers are not yet being widely used in linguistics is mainly that their presence on the academic scene is so recent that linguists for the most part have not had the opportunity to become acquainted with them. Moreover, the process of becoming acquainted has been impeded by various prevalent misconceptions regarding the range of applicability of the computer, the difficulty of learning how to use one, and the like. Perhaps the most widespread misconception is that the computer is primarily, if not entirely, a glorified mathematical calculator. Strangely enough, this misconception is even widespread among mathematicians who use computers. Since they use the machines only for purposes of numerical calculation, many of them have remained unaware of the numerous nonnumeric applications which are possible. Related to this