grid computing

"The cloud" - on-demand distributed computing power

Andrew Walkingshaw — Sat, 10 May 2008 16:25:00 -0700

Description

Simulation scientists are mostly limited by both the number, and the speed, of the computers available to them. Really large simulations need really serious computer resources, but simulations like that are pretty rare; so the resources for them have been concentrated in regional grids or national centres like the UK's HPCx.

Economically this makes a lot of sense, but there's a lot of overhead; for instance, compute time has to be bid for long in advance of when it might actually be used. In that light, commodity on-demand computing services like Amazon's EC2 begin to look promising as an alternative; they have even greater economies of scale than the national infrastructure services, can provide a scientist with more CPU power at essentially no notice, and often provide more flexibility in choice of operating system and software than a centrally-provided system can. At the moment, they don't scale to the massively parallel calculations that the national supercomputers specialize in, but sooner or later they'll be competitive even for those cases.

Computer & Information Science

Source

http://www.grid-support.ac.uk/content/view/239/157/
http://www.hpcx.ac.uk/
http://aws.amazon.com/ec2

Broadening Amateur Participation in Science

Alex Soojung-Kim Pang — Tue, 23 Oct 2007 11:10:30 -0700

Description

Interested amateurs are likely to have increased opportunities in the future to donate resources, time, or labor in support of scientific research, thanks largely to low-cost distributed computing.

The growth of peer-to-peer networking systems has created opportunities for amateurs to play a role in scientific research by donating computer time or labor. The pioneers in this arena are SETI@Home, Folding@Home, and other projects that invite people to load a piece of analytical software onto their computers. During periods of inactivity, the software downloads some data, analyses it, and then sends back the results. These programs enable those with computers to "donate" processor cycles to computationally intensive scientific or charitable activities.

It's important to remember the difference between:

Doing science on a personal level, and for the individual being involved in the science as a scientist. Advanced computer systems could help leverage individuals.

Exploiting distributed resources (e.g., SETI@Home) without the individual participating much themselves. Other examples are informed participation in medical developments (e.g., on the individual). In the future, people (and their houses, etc) will have lots of sensors, so possibilities here are substantial, especially for informing social policy (energy use, etc).

Gathering data, typically geographically specific data, or otherwise being a lab assistant, the individual devoting time and basic labour. Involving school children here, especially, can make them feel part of doing science, which will (hopefully) influence them for the rest of their lives.

SETI@Home, Folding@Home and other experiments have shown that amateurs can donate their time to analyse scientific data directly. The NASA Clickworkers system put volunteers through a simple training program to do routine analysis of Martian landscapes. The success of the system suggests that complex professional tasks done by highly trained and salaried individuals can be reorganized to tap a vast pool of tens of thousands of trained volunteers.

The strategy of Clickworkers and SETI@Home is to make science more accessible by making pieces of it very simple and by taking advantage of low-cost computing and communications. In the future, it is possible that more scientific research projects willdraw upon volunteered equipment or labour. In addition to distributed computing projects and efforts to mobilize volunteer observers, volunteers could be involved in gathering data using existing mobile communications or computing technologies -- for example, taking pictures of flora and fauna at specified times, or noting the GPS coordinates of certain objects.

Peer-to-peer and analytical computing projects have shown that it is possible to mobilize massive quantities of unused processing power or unskilled labour to do basic data analysis; such groups could be mobilized by advocacy and interest groups (e.g., supporters of breast cancer research or environmental causes) to create massive networks of volunteer labour. Expert knowledge that currently is underused in scientific research could be harnessed by custom-designed instruments with simple interfaces Finally, a new generation of sensor and smart dust technology could be used to make small instruments that volunteers carry with them, scatter about their environments, or leave in specific places, thus increasing scientists' mobility.

This will be enabled by:

Falling cost and increasing ubiquity of mobile communications and computing technologies
Growth of the open source movement
Establishment of the precedent of distributed computing projects in the 1990s and 2000s

Early indicators include:

"Proliferation of open-source, distributed computing and analysis projects such as Clickworkers and SETI@Home"

What to watch:

Volunteer projects are organised around popular issues like climate change and pollution.
NGOs and advocacy groups like Greenpeace or the World Wildlife Fund organise research projects for amateurs.

Signals

Supercomputing on Demand

Alex Soojung-Kim Pang — Tue, 23 Oct 2007 11:10:30 -0700

Description

New applications for supercomputing may develop over the next decade as large-scale supercomputing services become accessible over broadband terrestrial and wireless Internet networks by 2015.

Today, effective applications of supercomputing are mostly limited to industries such as petroleum and energy, aircraft and automotive design, and pharmaceuticals. Over time, these capabilities will migrate to mass markets as new applications in media, gaming, and ubiquitous computing will demand increasing cycles and require massive computing resources. The major computer and Internet companies are already recognising and tapping into the huge market opportunity this offers.

On-demand supercomputing is only one of many names for the idea of very high-performance computing programs linking supercomputer systems across broadband networks. This technology, currently under development in a whole family of research programs, is also known as grid computing, autonomic computing, adaptable computing, cluster computing, utility computing, and agile IT. The intention is to make computational power accessible in the same way that electricity is available from the electric grid -- users simply plug into it without worrying about where the power is coming from or how it got there. In this method of computing, if more computing power is required, spare cycles on other computers are used. This means that the power of supercomputing is accessible without the huge costs of supercomputing, and that CPU cycles that would otherwise be wasted are put to good use. So far, the fundamental building blocks of most research grids are commodity microprocessors linked into Linux clusters.

According to a DARPA study released July 16, 2005, only a fraction of available online high-performance computing resources are actually being used. The main reason for this is the difficulty and expense of programming new applications. But given the relentless progress of multicore and nanoscale processor design, the demand will increase for programmers to learn how to program massively parallel and threaded applications. By 2015 the programming obstacles to development could largely be solved.

On-demand supercomputing may increasingly be used for ordinary pervasive computing, sensor nets, speech recognition, language translation, image recognition, online games, and ubiquitous media. Additionally, industries will increasingly benefit from capabilities to casually use huge numerical models, very high-resolution simulations, and real-time interactive graphic models. For instance, media companies like George Lucas's Industrial Light and Magic, Steven Speilberg's Dreamworks, and Steve Jobs's Pixar Productions are already using massively parallel process to render movie graphics. Increasingly, media companies are sending jobs over the Internet to centralised "render farms" rather than maintaining their own computing resources. Those that do maintain their own render-farms can sell spare clock-cycles to others.

Interestingly, the idea of using networking to share computer power rather than just data is not new, it was the original role conceived for the Internet. Contrary to popular belief, the Internet was not originally conceived as a communication medium. Arpanet (the precursor to the Internet) was designed to give geographically distant researchers access computing resources because the only alternative, building computers in every institution, was, in the 1960s, too expensive to be considered. The idea of linking researchers through their computers only came later.

This will be enabled by:

Increasing supply of skilled programmers of massively parallel applications
Continued research that improves the ease of use of software development programs for massively parallel and threaded applications

Early indicators include:

Current offerings by IBM, HP, Sun, Oracle, SAS, and others of grid service extensions to existing ASP-hosted Web services
Google's use of a massively distributed load-balanced network of thousands of generic Linux computers to build one of the largest and fastest growing cluster-based services, offering anyone on the Web free storage of videos and unlimited mail storage, and offering professional users of a premium version of Google Earth a broadcast quality, real-time interactive, zoomable, navigable globe of the Earth
Current offerings by Yahoo, Microsoft, and AOL of services using massive distributed resources
Current offering by Apple's XGrid of instant distributed computing capabilities for any application
The emergence of commercial 'render-farms' selling services over the Internet to special effects houses and film studios
The development of grid computing
Increasing number of applications for distributed computing

What to watch:
Rate of consumer adoption of high-resolution graphic network applications like multiplayer games, Google Earth, Microsoft Virtural Earth, and Internet HDTV shows a steady increase.
Applications for interactive high-resolution graphics continue to expand in enterprises like pharmaceuticals and energy production.

Signals

More Accurate Modeling of Complex Economic Systems

Alex Soojung-Kim Pang — Tue, 23 Oct 2007 11:10:30 -0700

Description

Advances in simulation tools and behavioural analysis may facilitate innovation in economic research methods.

Within 10 to 20 years, agent-based modelling may facilitate a significant improvement in the scope and utility of economic models. Accurate simulations could potentially give economists new confidence in their conclusions. The difference this could make to economics can be compared with the impact of structural engineering on building. In the past, people designed buildings by intuition, experience, and guesswork. Today novel structures can be built with confidence because we have theories of structures and materials and we can model buildings with computers before they are built. Agent-based modelling could have a similar impact on economics.

Two factors are driving change in economics: accessibility of computational power and the adoption of behaviourist approaches. The (increasing) accessibility of computational power permits the accurate and rapid modelling of complex economies. The availability of computational power may, for a new generation of researchers, facilitate a shift in economics from deductive formalism to an applied mathematical approach based on simulations. In addition, the fundamental simplifying assumptions underlying current economic theory (greed, rationality, and equilibrium) are tending now to be replaced by new insights from behavioural studies of economic actors (firms, consumers, etc.) Drawing upon psychological experiments, behavioural approaches to economics may provide a far more accurate model of economic behaviour, revealing how variations in behaviour contribute to complexity in economic systems.

The impact of behavioural approaches to economic theory and computer power combined, may provoke a widespread shift in economic research methods from rational-actor models towards large simulations of complex economies inhabited by behaviourally-sophisticated agents. Cheap grid computing power will allow massive experiments that will begin to explain many of the seemingly random everyday trends in complex economic systems like the financial markets and international trade.

Implications:
* Ability to explore a far wider range of public policy alternatives than previously
* Creation of large new areas of inquiry for economists
* Potential fo expansion of economic theories and analytical techniques in other social sciences

Early Indicators:
* Awarding of recent Nobel prizes that lay the foundation for doctoral student interest behavioural economics
* Increase in the number of doctoral dissertations having to do with agent-based modeling techniques, and increase in the funding of research taking these approaches
* Work by a consortium of universities in a project known as NEW-TIES -- New and Emergent World models Through Individual, Evolutionary, and Social Learning -- on developing a large-scale and highly complex computer-based society as a means to understand social learning and behaviour

What to Watch:
* Faculty research in economics reflects increasing interest in agent-based models and behavioural analysis.
* Graduate economics curricula change to encompass agent-based models and behavioural analysis.
* Results from behavioural/ABM analysis are increasingly used in policy debates.

Parallels/Precedents:
* Complexity revolution in the "hard" sciences -- physics, biology, and computing

Enablers/Drivers:
* Increasing availability of cheap desktop computational power
* Continued development of high-level programming languages for building models
* Continuing or expanding support for interdisciplinary work between economists and computer scientists, mathematicians, and psychologists
* Neurotechnology for gathering data about emotional state of experimental subjects

Signals

New Dominance of Parallel Programming

Alex Soojung-Kim Pang — Tue, 23 Oct 2007 11:10:29 -0700

Description

Parallel programming -- programming for hundreds or thousands of concurrent independent processes or 'threads', may become increasingly important over the next decade as the result of developments in both hardware and software. Programming for small scale mobile and embedded devices may be an exception to this trend.

By 2015, parallel or concurrent programming may be the dominant mode of coding, rather than the niche technique it is today. A variety of new computing architectures will necessitate parallel programming:

Virtual computers -- ordinary computer integrated circuits with multiple computing cores within a single chip
Nanoscale computers -- nanoelectric, nanomechanical, nanobiological, and perhaps quantum computers
Grid, or cluster, computing over broadband networks

This future capability is increasingly indistinguishable from what we call supercomputing today. In the future, even mobile devices will likely have access to high-performance supercomputing over broadband networks. To benefit from parallel hardware, the software must provide enough concurrent operations to use all the hardware.

Today, the system software used on most contemporary supercomputers is a crude variant of UNIX with limited programs written in Fortran, C, and C++, augmented with a few language or library extensions for parallelism and with application libraries like the Titanium extensions to Java for higher performance computing. Even though some early software development tools exist, programming supercomputer software that supports massive parallelism requires special expertise to optimise performance at multiple levels below the operating system. Modern mainstream application programmers who are deeply knowledgeable about their application domains typically have little or no practical knowledge of using massive parallel processes to generate better results or user experiences. They usually employ high-level tools that mask the complexity of lower-level, massively parallel software and hardware processes. Indeed, application designers at large have not been educated in how to think about discretising computational tasks to take advantage of the improved computing machinery that is expected to be widely available starting around 2015.

Designers of parallel computing applications will need to be capable of determining the utility and cost of a solution based on expert judgments of factors other than time taken -- for instance, on accuracy or trustworthiness. According to a recent National Academy of Sciences study, 'Determining the trade-off among these factors is a critical task. The calculation depends on many things -- the algorithms that are used, the hardware and software platforms, the software that realizes the application and that communicates the results to users. . . . The design of the algorithms, the computing platform, and the software environment governs performance and sometimes the feasibility of getting a solution'."

This will be enabled by:

"Inherent parallelism of almost all new computer architectures
Demand for massively parallel software for secret government signal sensing and cryptographic applications, as well as for petroleum, automotive, and aircraft companies and pharmaceutical and biomedical start-ups"

Early indicators include:

"Announcement by Intel that it is investigating three fundamental types of processing capabilities necessary to deal with massive computing workloads: Recognition, Mining and Synthesis, or RMS
Increasing use by media companies such as George Lucas's Industrial Light and Magic, Steven Speilberg's Dreamworks, and Steve Jobs's Pixar Productions of massively parallel process to render movie graphics.
The growth of commercial "render-farms" selling services to media companies.
The growth of grid computing
Increasing number of distributed computing applications
Launching of pilot programs at the Centre for Parallel Computing at Massey University in New Zealand and the University of California at Berkeley to begin teaching design of massively parallel applications"

What to watch:

Photorealistic capabilities enabled by massively parallel computers migrate to interactive entertainment like games and other new high-resolution media.
Courseware in parallel processing is introduced at major educational institutions.