Infinity Supercritical LLC


Publication Title | Evaluating the Benefits of An Extended Memory Hierarchy for Parallel Streamline Algorithms

SDR Search Engine Series

Spinning Disc Reactor | Botanical Oil Organic Fluid Extraction search was updated real-time via Filemaker on:

Spinning Disc Reactor | Botanical Oil Organic Fluid Extraction | Return to Search List

Search Completed | Title | Evaluating the Benefits of An Extended Memory Hierarchy for Parallel Streamline Algorithms
Original File Name Searched: LBNL-camp.pdf | Google It | Yahoo | Bing



Page Number: 001
Previous Page View | Next Page View

Text | Evaluating the Benefits of An Extended Memory Hierarchy for Parallel Streamline Algorithms | 001



Evaluating the Benefits of An Extended Memory Hierarchy for Parallel Streamline Algorithms

David Camp∗ Lawrence Berkeley National Laboratory and Department of Computer Science, University of California, Davis

ABSTRACT

Hank Childs† Lawrence Berkeley National Laboratory and Department of Computer Science, University of California, Davis

Amit Chourasia‡ San Diego Supercomputer Center, University of California, San Diego

Christoph Garth§ Department of Computer Science, University of California, Davis

Kenneth I. Joy¶ Department of Computer Science, University of California, Davis

The increasing cost of achieving sufficient I/O bandwidth for high end supercomputers is leading to architectural evolutions in the I/O subsystem space. Currently popular designs create a staging area on each compute node for data output via solid state drives (SSDs), lo- cal hard drives, or both. In this paper, we investigate whether these extensions to the memory hierarchy, primarily intended for com- puter simulations that produce data, can also benefit visualization and analysis programs that consume data. Some algorithms, such as those that read the data only once and store the data in primary memory, can not draw obvious benefit from the presence of a deeper memory hierarchy. However, algorithms that read data repeatedly from disk are excellent candidates, since the repeated reads can be accelerated by caching the first read of a block on the new resources (i.e. SSDs or hard drives). We study such an algorithm, streamline computation, and quantify the benefits it can derive.

Index Terms: Programming Techniques [D.1.3]: Concurrent Programming—Parallel Programming, Computation by Abstract Devices [F.1.2]: Modes of Computation—Parallelism and Con- currency, Computer Graphics [I.3.3]: Picture/Image Generation— Display Algorithms

1 INTRODUCTION

As supercomputers get ever larger, the cost of achieving sufficient I/O bandwidth is, unsurprisingly, increasing. But supercomputing architects have been experimenting with a new approach to de- crease this cost. Where the typical approach has a simulation write data directly to a parallel file system (i.e. “spinning disk”), the new approach introduces a new participant, solid state drives (SSDs), and has the simulation write data to the SSDs instead. The simula- tion can then immediately resume, while, concurrently, the data is copied from the SSDs to the file system, shielding the simulation from slow parallel file system performance. Although the SSDs in- troduce a new cost, they lessen the importance of I/O bandwidth, allowing for the SSDs to be coupled with a slower (and less expen- sive) parallel file system, providing a cost reduction overall.

To applications, this I/O configuration appears to have two dis- tinct bandwidth characteristics. On write, the bandwidth appears to be good, since it is be accelerated by SSDs. On read, however, the bandwidth will be poor, since the reads are backed by a slower

∗e-mail: dcamp@lbl.gov †e-mail: hchilds@lbl.gov ‡e-mail: amit@sdsc.edu §e-mail: cgarth@ucdavis.edu ¶e-mail: kijoy@ucdavis.edu

IEEE Symposium on Large-Scale Data Analysis and Visualization October 23 - 24, Providence, Rhode Island, USA

©2011 IEEE

parallel file system and the presence of SSDs can not accelerate this activity.

I/O is often the slowest part of a visualization pipeline [10], hence suboptimal I/O read performance will result in poor overall visualization performance. However, in this paper, we ask whether SSDs can effectively increase I/O performance – and therefore vi- sualization performance – by treating them as an extended part of the memory hierarchy. While the first read of any block of data will remain slow, the SSDs can be used as a cache to store those blocks, considerably accelerating subsequent reads. Further, local hard drives are appearing increasingly commonly in the I/O sub- system and can similarly be used as an extension to the memory hierarchy in the same fashion as SSDs. We also study how these hard drives can accelerate I/O and visualization performance.

Although many paradigms for processing data do not read blocks of data repeatedly, streamline calculations do. Streamlines, or more generally integral curves, are one of the most illuminating techniques to obtain insight from simulations that involve vector fields and they serve as a cornerstone of visualization and anal- ysis across a variety of application domains. Drawing on an in- tuitive interpretation in terms of particle movement, they are an ideal tool to illustrate and describe a wide range of phenomena encountered in the study of scientific problems involving vector fields, such as transport and mixing in fluid flows. Moreover, they are used as building blocks for sophisticated visualization tech- niques (e.g., [14, 17, 19]), which typically require the calculation of large amounts of integral curves. Successful application of such techniques to large data must crucially leverage parallel computa- tional resources to achieve well-performing visualization.

Among visualization techniques in general, streamline-based ap- proaches are notoriously hard to parallelize in a distributed memory setting [23], because runtime characteristics are highly problem- and data-dependent. In terms of parallelization approach, stream- line computations may be parallelized over data, parallelized over streamlines, or some hybrid between the two. When parallelizing over streamlines (or, equivalently, over their seed points), particles are advected and blocks of data loaded dynamically based on the trajectory taken. This is exactly the data processing pattern that can benefit from an extended memory hierarchy and we study this approach here.

In this paper, we study the benefits a local disk – either SSD or local hard drive – can provide to accelerate a parallel streamline algorithm. We perform a variety of tests and present results to show what I/O benefits can be gained with the use of an SSD or local hard drive compared to a parallel file system.

2 RELATED WORK

2.1 Parallel Particle Advection

The parallel solution of streamline-based problems has been con- sidered in previous work using a multitude of differing approaches. Generally, both data set, represented as a number of disjoint blocks,

57

Image | Evaluating the Benefits of An Extended Memory Hierarchy for Parallel Streamline Algorithms



evaluating-benefits-an-extended-memory-hierarchy-parallel-streamline-algorithms
Infinity Supercritical Spinning Disc Reactor Botanical Oil Extractor | Oil Extract Using Infinity Supercritical SDR Extraction System | Organic Method of Oil Extraction - Organic Botanical Extraction System Uses Water As A Solvent - Go to website

Search Engine Contact: greg@infinitysupercritical.com