The top research stories of the week have been hand-selected from prominent journals and leading conference proceedings. Here’s another diverse set of items, including lessons learned from system failures; a cross-platform OpenCL implementation; the best memory to extract GPU’s potential; innovative ideas for next-generation interconnects; and the benefits of cloud storage to HPC applications.
Learning from Failure
A recent paper [PDF] authored by Charng-Da Lu, Computational Scientist at the Center for Computational Research at SUNY at Buffalo, investigates the important topic of HPC system failures. The research team presents 8-24 months of actual failure data generated by three HPC systems at the National Center for Supercomputing Applications (NCSA).
Lu explains the impetus for the research thusly: “Continuous availability of high performance computing (HPC) systems built from commodity components have become a primary concern as system size grows to thousands of processors. To design more reliable systems, a solid understanding of failure behavior of current systems is in need.”
Learning from mistakes is essential to progress, and Lu argues that failure data analysis of HPC systems has three main goals:
1. It highlights dependability bottlenecks and serves as a guideline for designing more reliable systems.
2. Real data can be used to drive numerical evaluation of performability models and simulations, which are an essential part of reliability engineering.
3. It can be applied to predict node availability, which is useful for resource characterization and scheduling.
The analysis shows that the three systems had an availability of between 98.7-99.8%. Lu finds that most outages were caused by software halts, while downtime per outage was highest in the case of hardware halts or scheduled maintenance. His team employed failure clustering analysis to identify several correlated failures.
Next >> Box Counting Algorithm on GPU
Box Counting Algorithm on GPU and multi-core CPU
In the prestigious Journal of Supercomputing, Jesús Jiménez and Juan Ruiz de Miras from the Department of Computer Science, University of Jaén in Spain, have authored a paper recounting their work with a cross-platform OpenCL implementation of the box-counting algorithm – one of the most popular methods for estimating the Fractal Dimension.
The Fractal Dimension, they explain, is an effective, but time-consuming image analysis method used in many disciplines, including the biomedical field, environmental science, materials science and computer graphics. When it comes to the analysis of 3D images, box counting proves especially slow-going.
“Unlike parallel programming models that strictly depend on the hardware type and manufacturer, like CUDA,” the team writes. “OpenCL allows us to provide an implementation suitable for execution on both GPUs and multi-core CPUs, whatever the hardware manufacturer.”
Drawing on the work of earlier research, the authors design an OpenCL algorithm that has been specifically optimized according the type of the target device. They claim average speedups of 7.46× and 4×, when executed on the GPU and the multicore CPU respectively, compared to single-threaded (sequential) CPU implementation.
Can PCM Benefit GPU?
A new technical report from the College of William & Mary Department of Computer Science examines the benefits of deploying phase change memory (PCM) in tandem with GPU systems.
The seven-member research team starts with the following premise:
“Recent years have seen a rapid adoption of Graphic Processing Units (GPU) for computing beyond graphics processing. As a massively parallel architecture, GPU has demonstrated appealing energy efficiency and tremendous throughput. However, the energy efficiency of current GPU systems is still far from meeting the requirement of extreme-scale computing.”
“Can PCM Benefit GPU?” – this is the question posed by the researchers and the title of their 11-page paper [PDF]. They point to recent studies that highlight PCM’s energy efficiency potential when teamed with CPU systems that have a modest level of parallelism. But would the same benefits apply for GPU-like massively parallel systems?
The authors claim that their work is the “first systematic investigation into this question.” They conclude that promise of PCM-based memory for increasing the energy-efficiency of parallel CPU-based systems did not hold true for GPU computing. In fact, the use of PCM in tandem with GPUs significant degraded energy-efficiency. The authors pointed to a “mismatch between those designs and the massive parallelism in GPU” and further note that repairing the mismatch requires “innovations in both hardware and software support.”
Ultimately their work reconciles a hybrid memory design with GPU massive parallelism for enhanced energy efficiency. It is this design that they say yields 15.6% and 40.1% energy saving on average compared to DRAM and PCM respectively, with a performance hit of less than 3.9%.
Next >> Interconnects for Exascale
Interconnects for Exascale
As the coming generation of supercomputers reaches into exaflop-class territory, the HPC community faces fundamental challenges to the way that such systems are designed and operated. One the biggest hurdles will be powering and cooling these mammoth machines. Optical interconnects could help alleviate some of these issues and thus have been proposed as a potential exascale enabler, but they are not without challenges themselves, especially in regards to manufacturability.
The feasibility of implementing chip-to-board interconnects for high-performance computing is discussed in a recent paper published in the Feb. 22, 2013 edition of Proceedings of SPIE. Written by a team of European researchers, the paper makes the case for integrating optical interconnect technologies into the module and chip level.
The researchers argue that “the introduction of optical links into High Performance Computing (HPC) could be an option to allow scaling the manufacturing technology to large volume manufacturing. This will drive the need for manufacturability of optical interconnects, giving rise to other challenges that add to the realization of this type of interconnection.”
The authors envision a solution that puts optical components on the module level, integrating optical chips, laser diodes or PIN diodes as components. They note the method is analogous to constructing a surface-mount device (SMD), which has its components mounted directly onto the surface of printed circuit boards. This new class of 3-dimensional optical link is symbolic of the “fundamental paradigm shifts” that will usher in the exaflop future.
Next >> Evaluating Cloud Storage for HPC
Evaluating Cloud Storage Services for Tightly-Coupled Applications
This week’s HPC cloud item comes from a team of researchers from INRIA and Argonne National Laboratory. Their work “Evaluating Cloud Storage Services for Tightly-Coupled Applications” was published as a chapter in Euro-Par 2012: Parallel Processing Workshops.
Noting that past HPC cloud research primarily focused on performance as a way to quantify the HPC capabilities of public and private clouds, the team sets out to address the topic of data storage as it relates to traditional HPC applications.
“Tightly-coupled applications are a common class of scientific HPC applications, which exhibit specific requirements previously addressed by supercomputers,” write the authors. They’re referring to the fact that tightly-coupled applications work best when paired with a custom-tuned parallel file system (PFS). And while virtual machines can be outfitted with any file system, including PFS, the setup introduces issues around data persistency.
The research team elect to test a cloud-based storage service, and they opt for an open source platform as opposed to Amazon. They select the Nimbus Cloud framework and its S3-compatible storage service, Cumulus.
The group runs several experiments using an atmospheric modeling application running in a private Nimbus cloud. The results show that the application is able to scale with the size of the data and the number of processes (up to 144 running in parallel), while storing 50 GB of output data on the Cumulus cloud storage service.