Sunday November 13, 2005

Hagan Rivers, of Two Rivers Consulting, gave a talk on Ajax and Web 2.0 at last week's meeting of Boston CHI (the Boston-area chapter of the ACM SIG on Computer-Human Interaction.)

She is an excellent speaker and I found the Ajax section particularly exciting and the Web 2.0 piece less so. Probably because I feel Ajax is the early embodiment of something important, while Web 2.0 is a squishier concept--more an attitude than anything else.

So, what did learn? First, Ajax is what is enabling some of the new style of web interfaces that have started popping up. Panning by clicking and dragging in Google Maps, for example. Netflix is starting to incorporate Ajax into its interface. Web interfaces are starting to act more like desktop interfaces.

Second, I learned that Ajax really isn't new technology: it's a new name applied to an aggregation of several technologies. At its most basic, it is Asynchronous Javascript and XML. See the slide deck (below) for a more precise list of the technologies involved. What Ajax enables is asynchonous communication between the client and server sides of a web application. Old, page-oriented style of interface: User mucks locally with forms, etc., and then presses Submit which then might cause the server to generate new html for the client. New, Ajax-oriented style of interface: The client-side Ajax engine and the server have the ability to cooperatively and constantly update the application's web interface without explicit actions from the user. It's much closer to the kind of interface interactions that have been available for a long time in the desktop world.

So why is it exciting? If Ajax merely helps web apps move closer to what's already available for apps on the desktop, what is the big deal? The big deal is that you can now see an inkling of a future in which desktop operating systems and applications don't matter anymore. The browser really begins to be capable of becoming the future desktop replacement. And, while Ajax might not be exactly the right solution in the end, you can see the shape of where we are headed.

A desktop born of open standards and based on open standards. Applications that live in the network--applications and data that are accessible from everywhere. Looking out a few years, I see a broad and promising vista (with a small 'v') where a good, solid, standards-compliant browser is all you'll need to access your applications.

Hagan's slides and supporting video clips are here.

This is the final set of customer talks from the Sun HPC Consortium meeting in Seattle. About 150 Sun customers and numerous Sun employees have spent long days this weekend sharing information about customer experiences and needs on the one hand and information about Sun products and product futures on the other. All of this in advance the Supercomputing '05 conference proper, which kicks off officially tomorrow.

Phil Williams, University of Nottingham, UK

Dr. Phil Williams, an EPSRC Advanced Research Fellow at the University of Nottingham gave a talk titled Science and Performance at Nottingham. His talk was paired with a presentation by Michael Rudgyard of Streamline Computing. I've combined by notes from these talks here.

Phil presented some background information on Nottingham, including the fact that they have campuses in Malaysia and China, in addition to their multiple locations in the UK.

In their recent, large cluster procurement, Nottingham was looking for a partner who could deliver the most equipment for their budget. In addition to delivering a large cluster, the partner needed also to be able to deliver smaller, scaled down "clones" of this system for use at multiple locations within the University.

Phil mentioned how surprised they were when some of the vendors responded to the procurement proposal by showing up at their required technology demonstration sessions with sales people who were completely unable to answer technical questions from Nottingham personnel. [Amazing that anyone serious about being in HPC would not know that these customers are technically very savvy and should be engaged by people who can speak to them about their issues at a technical level.]

Seven vendors responded to their procurement request and six were invited to bid. The seventh, whose initial proposal exceeded the stated Nottingham budget by a factor of 2.5X, was not invited to the next level. Three vendors were selected to submit best and final offers and Sun was selected in September.

The Jupiter system, which was delivered in conjunction with Streamline Computing, consists of 1024 dual-processor v20z (Opteron) systems interconnected with Gigabit ethernet. The full system, which comprised 19 full racks, was built in Sun's Linlithgow facility as were the 16 smaller, clone systems.

This system is now #109 on TOP500 list at 3.14 TFLOP with a 72% efficiency rating on LINPACK. It is currently Sun's #1 system on the list and is the most efficient TOP500 gigabit installation in the TOP500.

Greg S. Johnson, Texas Advanced Computing Center, USA

Our final customer speaker of the event was Greg Johnson from the Texas Advanced Computing Center which is part of the University of Texas at Austin. His talk was titled Visualization at TACC.

TACC's mission and passion is distributed visualization. In addition, they are distributed visualization partners in the Teragrid, which provides scalable high-end visualization resources to users across the US.

Greg outlined the goals of distributed visualization. The first is providing access to high performance visualization of a power that is well beyond that available at a typical researcher's desktop. Second is location transparency of resources. And third is an improved end-user experience.

The challenges of distributed visualization were as follows. First, latency over the wide area network. Second, delivering quality of service at the user interface. And, third, WAN bandwidth (1280 pixels by 1024 pixels by 12 bytes/pixel by 24 frames/sec is about 360 MB/s uncompressed--can do much better with compression technology.) distrib viz challenges: latency (WAN and GPU readback)

Greg then described the overall architecture of the Sun Terascale Visualization System, which is the cornerstone of TACC's distributed visualization strategy. It was developed in a collaborative effort between TACC and Sun.

The major components of the system are a Sun Fire 25K SMP with 64 UltraSPARC IV processors and 500 GB of memory. In addition, a large number of NVIDIA Quadro FX3000G cards, Myrinet, and 3DLabs U22 cards are used as well in associated systems.

The main takeaways from Greg were that modern networks permit the wide area routing of geometry and pixels at interactive rates and that the access models for visualization resources in the mainstream are shifting in the direction of those of HPC.

I attended the Developer Tools breakout session of Sun's HPC Consortium meeting this morning in Seattle. It's amazing to see the turnout here and in other sessions at 8am on rainy, foggy Sunday morning. In addition to a set of Sun presentations, we had several customer presentations, which I've summarized below.

Dieter an Mey, RWTH Aachen University, Germany

Dieter an Mey, who is well-known to many of us involved in HPC at Sun, gave a presentation titled HPC Program Development. Dieter works at RWTH Aachen University where he is heavily involved in the universitied HPC efforts. He's also very involved in the industry-wide OpenMP effort.

Dieter presented results of several parallel application development and tuning efforts. His slides were graphical and dense, so I will attempt only a short summary here.

The Center now supports four platforms: SPARC Solaris, Opteron Solaris, Opteron Linux, and Opteron Windows. Doing this requires more tools and more compilers. And more work for Aachen staff. But more opportunity to attract a wider variety of users.

Dieter presented some results of a porting effort moving an application from Windows/Intel to Windows Opteron. He was able to achieve a 6.2x improvement overall, which was quite good. However, he could not resist also trying a port to Solaris/Opteron, which he did using a variety of Sun's tools, including his favorite performance analysis tool, Sun Analyzer. All in, all done, he achieved a speedup of 14x on Solaris Opteron. Nice.

Allen Malony, University of Oregon, USA

Professor Allen Malony gave a talk entitled TAU at Work about the TAU Parallel Performance System which has been developed over the past 14+ years. I highly recommend visiting the TAU site via the previous link as I can't possibly due this system justice here.

TAU is essentially a performance analysis framework for high performance computing. It is portable and open source. It supports multiple languages (C, F77, F90, C++, Java, etc.), multiple programming paradigms (multi-threading, message passing, hybrid models, etc.)

It supports performance instrumentation, measurement, and analysis--including scalable visualization in support of performance analysis. Instrumentation can be done at the source level (either manually or automatically), at the object level (via pre-instrumented libraries), or at the executable level via dynamic instrumentation.

TAU has been available previously on Solaris/SPARC and now will soon be available on Solaris 10 for Opteron. Cool.

Oscar Hernandez, University of Houston, USA

Oscar Hernandez, a graduate student at the University of Houston, spoke about the OpenUH project, which is creating an open source, portable OpenMP compiler suite that supports research into OpenMP language features, teaching, and exploration of tools for use with OpenMP.

Open64-UH is based on the Open64 compiler that was released to open source by SGI. It supports C, C++, Fortran77 and Fortran90.

Awhile back, I blogged about my experiences putting Solaris 10 on my laptop. I'm reporting in again, because I recently installed Nevada (which might eventually be called Solaris 11) along with a couple of extremely useful scripts that will greatly improve your Solaris laptop experience. The cool thing is that these scripts (and Nevada itself) are available on the Open Solaris website, so it's okay for me to blog about them.

The two scripts are inetmenu and frkit. While both are useful, inetmenu is the true godsend in that it automates the act of attaching to a network and makes it very easy to deal with the variety of networking environments a mobile user encounters: it can automatically configure using a static IP, DHCP with NIS, and DHCP without NIS. And it knows about wireless as well (if configured).

Check out the Laptop Community on the OpenSolaris site for more information on mobile Solaris. You'll find frkit, inetmenu, wireless, and other information there. Go wild and have fun!

By the way, I'm blogging this on my Nevada-enabled Tecra M2 laptop at Sun's HPC Consortium meeting in Seattle, connected via local wireless and VPN'ed into Sun. Smooth as silk.