Journal of the Mechanics and Physics of Solids, 60(5):1002-1019, 2012. DOI 10.1016/j.jmps.2011.12.00

Journal of the Mechanics and Physics of Solids, 60(5):983-1001, 2012. DOI 10.1016/j.jmps.2011.12.001.

Presented is a speculative server blade architecture called a FlashBlade that combines 100x I/O performance in both latency and bandwidth with balanced computing. The blade consists of a standard multi-core CPU with attached DRAM. It uses a fast interconnect, such as Intel’s QuickPath, to communicate with a FPGA router called the X1. This router handles traffic to the “C1 complexes” and off-blade. Each C1 complex is a System on a Chip with Package on Package DRAM, connected to local flash memory. There are numerous complexes, giving tremendous I/O performance and computational balance. A large design space of parameters such as flash size, number of complexes, and link bandwidth between each C1 and the X1 is available for power and performance optimization. A single blade server constructed from these blades, just 12.25 inches high and drawing about 10 KW, could support a few hundred thousand basic web searches a second on 1 billion pages. It could also provide triple store performance 100x greater than achievable now for datasets of 6 TB and scales to petabyte datasets although at somewhat reduced performance; with numerous applications to defense, commerce, and science.

(PDF)

Seminar given at “IST Lunch Bunch”
Tuesday November 2, 2010
12:00 PM
105 Annenberg

Abstract:
Moore’s law works for semiconductor-based detectors and there is an increasing flood of data being generated in astronomy, high energy physics, biology, and other sciences. Computation is essential for both (1) making predictions from theory and (2) the analysis of data from experiments. Is there a way to architecturally balance both needs in a high performance computing (HPC) system?

HPC systems are typically constructed from easily available parts, just organized differently and scaled to process extreme workloads. The most power efficient petascale machine as of 2010 is Roadrunner at the Los Alamos National Laboratory. Another interesting machine is the Apple iPad which uses very low power System on a Chip/Package on Package technologies. This talk explores the question of “what happens when you cross Roadrunner with iPads”? The result is a high level of integration between computation and storage on a single server blade, called a Flashblade, with 100x-1,000x performance improvements for some data-centric applications. The Flashblade architecture, expected performance, programming, and scaling with advancing technology are discussed.

(PDF)

Mark Stalzer

High performance computing systems are typically constructed from easily available parts, just organized differently and scaled to process extreme workloads. The most power efficient petascale machine as of 2010 is Roadrunner at the Los Alamos National Laboratory. Another interesting machine is the Apple iPad which uses very low power System on a Chip/Package on Package technologies. This talk explores the question of “what happens when you cross Roadrunner with iPads”?

(PDF)

Moderator:

• Mark Stalzer, CACR

Panelists:

• Thomas Sterling, Louisiana State University
• Allan Snavely, San Diego Supercomputer Center
• Stephen Poole, Oak Ridge National Laboratory
• William Camp, Intel Corporation

Abstract:
Over the next few years there are two boundary conditions that should constrain computer systems architecture: commodity components and applications performance. Yet, these two seem strangely disconnected. Perhaps we need some human optimization, as opposed to repeated use of Moore’s Law. Our panelists have been given a set of standard components that are on announced vendor roadmaps. They also each get to make one mystery component of no more complexity than a commercially available FPGA. The applications are HPL for linear algebra, Map-Reduce for databases, and a sequence matching algorithm for biology. The panelists have 10 minutes to disclose their systems architecture and mystery component, and estimate performance for the three applications at 1MW of power.

(PDF)

Abstract
The purpose of computational science and engineering is to use computers to accelerate scientific discovery and engineering design. This has been a major driver in the development of high performance computing. Yet, something is missing. In this talk I suggest a more systems engineering viewpoint of CSE where all parts are considered: sensors, computers, algorithms, and more formally connecting simulations to experiment. In this viewpoint, the question changes: from how fast does the computer go, to how much science can be done for fixed resources. Examples are given from work at CACR.

Slides (5MB PDF)

Sean Mauch and Mark Stalzer

(Abstract PDF) (Link to article)