Random thoughts on Google maps

As I headed out on the road for SearchSOA.com, I needed driving directions. Needed to find the bucolic burgh of, well let’s call it Medfordshire, Massachusetts. Where it is does not really matter.

To get driving directions, the evening previous to this excursion, I employed the great compute cloud known as Google. I asked this learned hand for instructions, and printed out a map that told me to proceed from my home in Boston’s Mission Hill, and take Route 95 North, etc.

Everything is going well. It is the next morning. Time to go. Alas, I’d left my Google Map print out back at the office. No problem, of course, I go online and get directions again.

After heading to Route 95 I pull out the directions and discover Google has computed a completely different set of plans to take me from and to the same end points. I put on my human hat and successfully made the trip. No problem. But it did get me to thinking about the compute cloud.

Clearly, somewhat random results work in some cases. But, a lot of enterprise computing must provide far more solid results. Read the rest of the story.

Cray Day! Make way for the PetaScale Jaguar

Where's Seymor?

It is an upgrade. But way up. How about a quadrillion mathematical calculations per second? Sounds prettier  than ‘1.64 petaflops, n’cest pa?  Yes, the folks at Oak Ridge have a hot-rod CPU. The makers call it Jaguar and mark it the world’s first petaflop system dedicated to open research. So apparently there’s more flops in the dark.

 

The upgrade at DOE’s Oak Ridge National Leadership Computing Facility represents a major milestone in a four-year project, begun in 2004 when DOE’s Office of Science launched a sustained effort to upgrade supercomputing capabilities for unclassified research at DOE’s complex of national laboratories.

Jaguar uses over 45,000 quad-core Opteron processors and features 362 terabytes of memory and a 10-petabyte file system. The machine has 578 terabytes per second of memory bandwidth and input/output (I/O) bandwidth of 284 gigabytes per second. 

Cray is on the move on other fronts as well. They have just announced  availability of NVIDIA Tesla C1060 GPU Computing processors in the new Cray CX1 line. Fits into a regular office without humongous refrigeration. Each Tesla processor has hundreds of processor cores that deliver nearly one teraflop of peak computing performance, we are told. It’s the revenge of the GPUs, one hopes.

http://www.energy.gov/news/6712.htm

A Noble Award for IBM magneto-wizard Parkin

In case you missed it, IBM fellow Stuart Parkin was awarded the Daniel E. Noble Award for fundamental contributions to the development of magneto-resistive devices for nonvolatile RAM. These advances greatly enhanced the capabilities of modern disk drives. In recent years Parkin has been prominent in work that led to Racetrack Memory, which uses electron spin, rather than the charge, to create electronic devices. Parkin received the award along with Jim Daughton and Saied Tehrani. The award was announced at the IEEE International Magnetics Conference in Madrid.

Grid tackles patient outcomes

IBM and several universities have banded together to form a collaborative research effort to create diagnostic tools for predicting cancer patients’ response to treatment.

 

The primary objective of the center is to develop pattern recognition algorithms that can compare data in digitally archived cancer specimens, radiology images and proteomic and genomic data. As such, this effort recalls other IBM-Academic Grid efforts that try to tap data bases of patient outcomes and tests.

 

In the effort, Reuters will address computational and distributed computing issues at the system and application levels. CINJ and Rutgers will develop suitable machine learning, image processing and pattern recognition methods. The NSF will work on autonomic systems and applications. Among other things, IBM is going to provide a pretty psychedelic Grid.

 

 IBM and University Researchers to Develop Research Tools to Improve Cancer Patient Outcomes – IBM.com

IBM has silicon Mach-Zehnder electro-optic modulator

IBM continues to focus on ways of making multi-cores chips shuttle info between cores. On Dec 6, the company discussed using pulses of light through silicon, instead of electrical signals on wires. A silicon Mach-Zehnder electro-optic modulator does the work of converting electrical signals into pulses of light. The IBM modulator is 100 to 1,000 times smaller in size compared to previously demonstrated modulators of its kind.

Central to the modulator are silicon nanophotonic waveguides that control the flow of light on a silicon chip. The waveguides are made of tiny silicon strips. Strong confinement of light allows the IBM modulator to be dramatically scaled down in size.
 
In April IBM announced success with 3-D chip stacking that eliminated the need for metal wires that connect standard 2-D chips together. The April 3-D demo instead used through-silicon vias etched through a silicon wafer and filled with metal.

On-chip silicon nanophotonics work seems to continue to confirm IBM’s central role in semiconductor advancement. An IMB researcher said nanophotonics can do for computers what fiber optic networks did for the Internet. – Jack Vaughan

http://www-03.ibm.com/press/us/en/pressrelease/22769.wss

Multi to the core

In looking at the merging dual-core and upcoming multicore processor firmaments for American Scientist, Brian Hayes goes over the prehistory. That is..the Inmos Transputer and its beloved Occam language compatriot, Danny Hillis and the Connection Machine, and David Gelerntner and the Linda OS. He mentions that these ventures were flattened by the general onslaught of commodity products based on clusters, racks, and Ethernet. Of course the demise of the Soviet Union had a role too.

I have this recollection of Hillis et al suddenly trying to make a living doing filters on American Express transactions – tough slogging after the original Evil Empire packed up.

What are the pivotal issues looking forward? Hayes writes that writing in the mode of parallelsim is just hard for humans. Vendors of hardware and software seldom are so genuine. Super cooled computers, not too practical on desktops, may be the alternative, he says. But don’t hold your frozen breath.

He pens: “Writing correct concurrent programs
is not impossible or beyond human
abilities, but parallelism does seem to
make extreme demands on mental discipline.
The root of the difficulty is nondeterminism:
Running the same set of
programs on the same set of inputs can
yield different results depending on the
exact timing of events. This is disconcerting
if your approach to programming
is to try to think like a computer.”

The mind of the programming humanoid tends to be single threaded. Or, as Hayes writes: You may be able to walk and chew gum at the same time, but it’s hard to think two thoughts at once.

Google seems to have had some success with parallelism – in the age of clusters, and Hayes mentions a paper describing that success. Fairly recently none other than IBM saddled up to Google to address the threading conundrum.

Parallelism to date has been shielded from the everyday programmer by compiler and OS adepts. It has been writ that the massively parallel demon cannot be hidden from the everyday programmer people too much longer. -Jack Vaughan

Multicore chips could bring about the biggest change in computing since the microprocessor – AM SCI

Google and clusters – Labs.GoogleMult