The multi-touch enabled display is a composite of 28 projectors with a total resolution of 7168 x 3072 pixels. It supports both physical and acoustic gestures to pan and zoom data sets of arbitrary size. It is driven by a 31-node ROCKs v4.x visualization cluster. Each node is a Dell Precision 370 workstations (Intel Pentium 4 EM64T at 3.2 GHz, 2 GB RAM, HyperThreading, NVIDIA Quadro FX 3400 with 256 MB VRAM on a PCI Express x16 bus).

Given the age of the ROCKs distro and the supporting compute hardware, this display must have been in place for a number of years.

(See also: Navigating 13.3 gigapixels on a 22 megapixel display wall)
(See also: YouTube: University of Tromsø wall Demonstration)
(See also: ROCKs Clusters)

Despite several rendering issues on the SlideShare site, the presentation is available here:

(See also: SlideShare: Cloud Computing Services and the Data Center)

The following weekend an interesting article about data centers appeared in the New York Times magazine:

(See also: NYtimes Magazine: Data Center Overload)
(See also: NYtimes Magazine: Data Center Overload slides)

According to the NEBULA website:

The fully-integrated nature of the NEBULA components provides for extremely rapid development of policy-compliant and secure web applications, fosters and encourages code reuse, and improves the coherence and cohesiveness of NASA’s collaborative web applications. It is used for Education and Public Outreach, for collaboration and public input, and also for mission support.

NEBULA extends the Software-as-a-Service to the realm of Platform-as-a-Service and Infrastructure-as-a-Service. In the process, slaying several classic conundrums of computational collaboration.

I wish them the best of luck.

(See also: NEBULA site)
(See also: InformationWeek: NASA Launches Nebula Compute Cloud)
(See also: Open Eucalyptus project)
(See also: Eucalyptus Cloud Computing presentation)

It’s been an interesting evening while Stephen Wolfram launches his new computational knowledge engine ‘WolframAlpha‘. There were some fits and starts, but I have been able to ask it some interesting questions.

The approach reminds me of Thinking Machines Wide Area Information Server (WAIS) and Gopher from the late 1980′s. I know that Stephen Wolfram worked at Thinking Machines and I don’t know if he was involved with the WAIS project, but it certainly was a fundamental influence on WolframAlpha. WAIS was ultimately sold to AOL in 1995, just as the World Wide Web was forming.

WolframAlpha is easily stumped. But then you ask it a question that fans out into an amazing array of results from wide and varied data sources. TEDChris has side by side comparison of seven queries given to WolframAlpha and Google. A helpful illustration of the differences between the two philosophies. When the answer isn’t a precise number, WolframAlpha will try to reduce the question to something it can answer precisely. If the answer is a precise or computed number WolframAlpha can produce an elegant and concise response, though much of the supporting data appears to be older sources than those revealed in similar Google searches. While powerful in certain domains (such as math, chemistry, census data), the result is a service that may produce what you need or nothing useful at all. Here are some funny queries of interest:

There is a certain level of hubris in the idea all knowledge can be contained, maintained, and computed centrally. WolframAlpha is a ‘come to the mountain’ experience. In contrast, Google’s shotgun response relies on the distributed nature of the internet, counting and weighing the edges between ideas, often responding with a myriad of links relying on the user to be the final filter.

Both systems have a place in my toolbox.

(See also: WolframAlpha: query interface)
(See also: Introducing WolframAlpha)
(See also: WAIS: Wide Area Information Server)
(See also: TEDChris: WolframAlpha vs Google)
(See also: TechCrunch: Putting Wolfram Alpha To The Test: Not Super-Impressed)

There is now a Green500.org list ranking machines on the TOP500.org list based on energy efficiency, in FLOPS per Watt. LINPACK makes use of the BLAS (Basic Linear Algebra Subprograms) libraries to perform vector and matrix operations. The three benchmarks that comprise LINPACK were designed for use on vector processors in the 1970s and early 1980s.

If LINPACK is representative of the types of calculations and loads produced by your application suites, then Green500.org ranking may be a valid measure of performance efficiency. If LINPACK loads are not representative of your work, it would be interesting to explore a metric more appropriate to modern hardware and application loads. Perhaps LAPACK-based FLOPS over Watts metric?

(See also: TOP500.org)
(See also: Green500.org)

(See also: LINPACK: Fortran subroutines that analyze and solve linear equations.)
(See also: LINPACK: latest report)
(See also: LAPACK: Linear Algebra PACKage)
(See also: BLAS: Basic Linear Algebra Subprograms)

(See also: EnergyStar.gov)
(See also: GreenerComputing.com)

(See also: IEEE: Green Supercomputing Comes of Age)
(See also: IEEE: The Green500 List – Encouraging Sustainable Supercomputing)

(See also: Honey, I Shrunk the Beowulf)

Features of the Silicon Graphics Molecule concept include:

• High concurrency with 20,000 threads of execution — 40 times more than a single rack x86 cluster system
  
• High throughput with 15TB/sec of memory bandwidth per rack — over 20 times faster than a single rack x86 cluster system
  
• Greater balance with up to three times the memory bandwidth/OPS compared to current x86 CPUs
  
• High performance with approximately 3.5 times the computational performance per rack
  
• Greener with low-watt consumer CPUs and low-power memory that deliver 7 times better memory bandwidth/watt
  
• Innovative Silicon Graphics Kelvin cooling technology, which enables denser packaging by stabilizing thermal operations in densely configured solutions
  
• Operating environment flexibility, capable of running industry-standard Linux® implementations, with Microsoft® Windows® variants on some configurations
  

Classic problems appear to be solved. Speeds and feeds, heat and power. Not sure how they solve memory bandwidth bottlenecks at this density. Likely a NUMA (CC-NUMA?) architecture, but that would likely force memory down to the already improbably crowded processing modules. Cooling will be critical in a system such as this. Silicon Graphics® Kelvin™ cooling technology professes to solve this problem.

Granted, this is a prototype and may never be built for sale. It may never be built at all. I remain skeptical. I wish I were at SuperComputing 08 in Austin, TX this week to see it up close. If you are, please leave a comment.

(See also: SGI: a Glimpse of the Future)
(See also: Wired: SGI Creates 10,000-core Concept Computer)

Recent expansion has doubled performance of the Jaguar computer at ORNL to 119 teraflops of peak performance (119 trillion mathematical calculations per second). It is now the most powerful open scientific computing system in the world.

Doubling the size of the system involved adding involving 124 cabinets. Jaguar uses over 45,000 of the latest quad-core Opteron processors from AMD and features 362 terabytes of memory and a 10-petabyte file system. The system has a rating of 119 teraflops of peak performance (a teraflop is 10¹²).

When I worked at Cray, the metric was simple multiples of sustained gigaflops. A gigaflop is one billion (10⁹) floating-point operations per second.

The upgraded Jaguar will undergo rigorous acceptance testing through December before transitioning to production in early 2009.

(See also: ORNL: Cray supercomputer doubles performance)
(See also: DOE’s ORNL supercomputer now world’s fastest for open science)

Today Cray introduced a 62lb desk-side supercomputer chassis, sporting either Windows HPC Server 2008 or Red Hat Enterprise Linux. Both operating systems come with a web-based management interface. Future support will include SuSE. The unit is also available in a 7U rack-mount weighing 136lbs, fully loaded.

The CX1 is Intel Xeon powered with up to 16 dual and/or quad core processors. Up to 8 nodes with up to 64 GB of RAM per node. Featuring compute, storage, and visualization blades. The CX1 runs on power from a standard wall socket (20amp/110/220v) and active noise canceling allows it to blend into the office environment.

U.S. list prices start at $25,000 continue over $60,000.
The cool bit is website lets you configure your own.

Has anyone seen one of these in the wild? Perhaps you need someone to benchmark some life science codes?

(See also: cray.com:products:CX1)
(See also: cray.com:CX1 demo)
(See also: cray.com:CX1:resources)
(See also: cray.com:press release:CX1)
(See also: Cray:blog)
(See also: Engadget:CX1)

In the column of things that are not cloud computing, the new Globus toolkit was released today. Highlights of this release include:

Persistent HTTP/S connection support in Java WS Core
Dynamic deployment support in Java WS Core
JBOSS 4.0.x support in Java WS Core
An implementation of WS-ServiceGroup added to C WS Core
C command-line tools for WSRF operations
Support for GetResourceProperties and QueryResourceProperties in the Delegation Service
Added support for the OGSA-AuthZ Authorization Service to CAS
Server-side attribute-based authorization framework enhancements
Support for a pluggable Policy Decision Point (PDP) designed to minimize common authorization errors
Enhanced security descriptor framework
A Web service interface for the Replica Location Service (RLS)
Improved support for multiple TriggerRules in the Trigger Service
Improved configuration interface for the Trigger Service
Java API to assist in creating resource properties from external information sources
A new resource manager (RM) adapter API in GRAM4

One item of note is the web service interface for the Replica Location Service (RLS).

With the advent of cloud computing models I’m increasingly frustrated by the direction of efforts like this. I understand the need for security models and tools for managing and processing across geographically distant topologies and authority domains. I understand that ‘grids’ and ‘cloud services’ are quite different ideas. I’m not convinced anyone would choose to use a grid, specifically the Globus toolkit, unless it was mandated by their research community and/or funding agency. (See caBIG.) That and the fundamental reliance of a high-performance distributed processing environment on JAVA is baffling. The overhead of this environment is immense and doesn’t appear to be slowing.

The threshold is too high. It simply isn’t worth the effort or supporting infrastructure to deploy a simple set of services for most research communities. RESTful services solve most of these problems with familiar and supported protocols, without the overhead.

(See also: Globus Toolkit 4.2 Released)
(See also: GLOBUS: release notes)
(See also: GLOBUS: software(See also: GLOBUS: documentation))

(See also: caBIG)

(See also: Dissertation: Architectural Styles and the Design of Network-based Software Architectures)

• If you peel back the label and it says “Grid” or “OGSA” underneath… it’s not a cloud.
• If there is a consultant in the room… it’s not a cloud.

That said, Cloud Camp is in London next week (July 16) and I’d dearly love to go. I’ve no doubt Werner will speak about Amazon web services and James Governor from Redmonk will be there to report.

Doesn’t cloud computing generally boil down to a conversation with your carriers? In my mind that’s the real question. How to you feel about trusting your enterprise to the cable guy? Perhaps not today, but eventually you will.

(See also: Monkchips: 15 Ways to Tell It’s Not Cloud Computing)

(See also: Gigaom: 10 Reasons Enterprises aren’t ready to trust the Cloud)

(See also: zdnet: How to deploy the cloud of your choice)

(See also: Cloudcamp: London)

(See also: Amazon Web Services)
(See also: BLOG: Werner Vogels)

(See also: Redmonk: Analysis for the people, by the people)
(See also: BLOG: James Governor)