Tara M. Madhyastha

Scalable Performance Environments for Parallel Systems

As parallel systems expand in size and complexity, the absence of performance tools for these parallel systems exacerbates the already difficult problems of application program and system software performance tuning. Moreover, given the pace of technological change, we can no longer afford to develop ad hoc, one-of-a-kind performance instrumentation software; we need scalable, portable performance analysis tools. We describe an environment prototype based on the lessons learned from two previous generations of performance data analysis software. Our environment prototype contains a set of performance data transformation modules that can be interconnected in user-specified ways. It is the responsibility of the environment infrastructure to hide details of module interconnection and data sharing. The environment is written in C++ with the graphical displays based on X windows and the Motif toolkit. It allows users to interconnect and configure modules graphically to form an acyclic, directed data analysis graph. Performance trace data are represented in a self-documenting stream format that includes internal definitions of data types, sizes, and names. The environment prototype supports the use of head-mounted displays and sonic data presentation in addition to the traditional use of visual techniques.

Input/output access pattern classification using hidden Markov models

Input/output performance on current parallel file systems is sensitive to a good match of application access pattern to file system capabilities. Automaticiuput/output access classification can determine application access patterns at execution time, guiding adaptive file system policies. In this paper we examine a new method for access pattern classification that uses hidden Markov models, trained on access patterns from previous executions, to create a probabilistic model of input/output accesses. We compare this approach to a neural network classification &n-rework, presenting performance results from parallel and sequential benchmarks and applications.

Learning to classify parallel input/output access patterns

Input/output performance on current parallel file systems is sensitive to a good match of application access patterns to file system capabilities. Automatic input/output access pattern classification can determine application access patterns at execution time, guiding adaptive file system policies. In this paper, we examine and compare two novel input/output access pattern classification methods based on learning algorithms. The first approach uses a feedforward neural network previously trained on access pattern benchmarks to generate qualitative classifications. The second approach uses hidden Markov models trained on access patterns from previous executions to create a probabilistic model of input/output accesses. In a parallel application, access patterns can be recognized at the level of each local thread or as the global interleaving of all application threads. Classification of patterns at both levels is important for parallel file system performance; we propose a method for forming global classifications from local classifications. We present results from parallel and sequential benchmarks and applications that demonstrate the viability of this approach.

Data sonification: do you see what I hear?

The authors assert that, despite great strides in developing the graphical dimension of user interfaces, the auditory dimension has been neglected. They offer a tool for using sound to complement visual cues when working with complex data. The Porsonify toolkit creates sonifications that can be easily integrated with existing visualization systems. Porsonify provides a platform to experimentally map data to sound. To provide portability, Porsonify relies on a network interface to sound devices and descriptions of abstract sound devices. By manipulating abstract sound devices, Porsonify allows access to unique, device-specific features, while hiding their complexities within the higher level concept of a sonification. This structure makes it easy to create sonifications, both in isolation and with accompanying visualizations. >

Intelligent, adaptive file system policy selection

Traditionally, maximizing input/output performance has required tailoring application input/output patterns to the idiosyncrasies of specific input/output systems. The authors show that one can achieve high application input/output performance via a low overhead input/output system that automatically recognizes file access patterns and adaptively modifies system policies to match application requirements. This approach reduces the application developers input/output optimization effort by isolating input/output optimization decisions within a retargetable file system infrastructure. To validate these claims, they have built a lightweight file system policy testbed that uses a trained learning mechanism to recognize access patterns. The file system then uses these access pattern classifications to select appropriate caching strategies, dynamically adapting file system policies to changing input/output demands throughout application execution. The experimental data show dramatic speedups on both benchmarks and input/output intensive scientific applications.

Physical Modeling of Probe-Based Storage

Magnetic disks may be reaching physical performance limits due to the superparamagnetic effect. To close the performance gap between processors and storage, researchers are exploring a variety of new storage technologies [17]. Among these new technologies, probe-based micro-electrical mechanical systems (MEMS) magnetic storage arrays are attractive [3]. Probe-based storage is dense and highly parallel. It uses rectilinear motion in contrast to rotating media. Commercial devices are expected within the next several years. The wide range of possible architectures and the unique performance characteristics of probe-based storage require that standard file system algorithms for disks, including scheduling and layout, must be revisited to determine their efficiency domain. Because these devices do not yet exist, analysis of system performance depends on simulation models. At this early stage of development, models that bridge the gap between the physics of the device and its performance characteristics can provide important feedback to both hardware and software designers. This paper compares results from three models of probe-based storage that convey successively more accurate descriptions of the underlying physics. We conclude that the physical accuracy of the model has a significant impact on the predicted performance under real workloads.

The relevance of long-range dependence in disk traffic and implications for trace synthesis

Accurate disk workloads are crucial for storage systems design, but I/O traces are difficult to obtain, unwieldy to work with, and unparameterizable. I/O traces are often bursty and difficult to characterize. Although good models of I/O workloads would be extremely useful, such bursty traces cannot accurately be modeled using exponential or Poisson arrival times. Much experimental evidence suggests that I/O traces are self-similar, which researchers have hoped might help to model bursty traces. In this paper, we show that self-similarity at large time scales does not significantly affect disk behavior with respect to response times. This allows us to generate synthetic arrival patterns at relatively small time scales, improving the accuracy of trace generation. The relative error of our method, with input parameters suitable for the workload, ranges from approximately 8% to 12%.

Exploiting Global Input Output Access Pattern Classification

Parallel input/output systems attempt to alleviate the performance bottleneck that affects many input/output intensive applications. In such systems, an understanding of the application access pattern, especially how requests from multiple processors for different file regions are logically related, is important for optimizing file system performance. We propose a method for automatically classifying these global access patterns and using these global classifications to select and tune file system policies to improve input/output performance. We demonstrate this approach on benchmarks and scientific applications using global classification to automatically select appropriate underlying Intel PFS input/output modes and server buffering strategies.

Workload based optimization of probe-based storage

The performance gap between microprocessors and secondary storage is still a limitation in todays systems. Academia and industry are developing new technologies to overcome this gap, such as improved read-write head technology and higher storage densities. One promising new technology is probe-based storage[1]. Characteristics of probe-based storage include small size, high density, high parallelism, low power consumption, and rectilinear motion. We have created a probe-based storage simulation model, configurable to different design points, and identify its sensitivity to various parameters.

I/O, performance analysis, and performance data immersion

A large and important class of national challenge applications are irregular, with complex, data dependent execution behavior, and dynamic, with time varying resource demands. We believe the solution to the performance optimization conundrum is integration of dynamic performance instrumentation and on-the-fly performance data reduction with configurable, malleable resource management algorithms, and a real-time adaptive control mechanism that automatically chooses and configures resource management algorithms based on application request patterns and observed system performance. Within the context of parallel input/output optimization, we describe the components of such a closed-loop control system based on the Pablo performance analysis environment, a portable parallel file system (PPFS), and virtual environments for study of dynamic performance data and interactive control of file system policies.