Do you find the following attention-catching question fascinating?

The XMT vision was introduced in 1997. The following recent highlights suggest that the world is getting much closer to that vision:
- Fueled by six decades of unrelenting exponential progress in CPU clock speed, the mainstream computer architecture industry focused exclusively on optimizing performance using standard serial software. However, rather than having 10 GHz CPUs today, as would have been expected based on past progress, high-end CPU speeds are at around 4GHz, more or less where they were in mid-2003. While the popular press has mostly missed this rather dramatic lack of progress, it is widely recognized among experts that this is not a coincidence and the old paradigm is indeed collapsing: serial computing is finally running out of steam. The recognition that business survival for that industry depends upon new parallel computer languages and parallel architectures is just starting to sink in.
- SUN announced its new Niagara architecture with 32 processors on chip.
- Intel has announced that all their future processor chips will have multiple processors on the chip. See also a recently proposed framework for 2005-2015 under: Platform 2015: Intel Processor and Platform Evolution for the Next Decade and note the trend towards parallel applications in the presentation by Justin Rattner, Intel.
- IBM, Sony and Toshiba announced their joint Cell chip multiprocessor. (Sony announced its deployment in the Sony Play Station 3.)
- In a 2005 article entitled The Free Lunch Is Over: A Fundamental Turn Toward Concurrency in Software, Herb Sutter, Microsoft, reviews how hardware vendors have run out of room with most of their traditional approaches to boosting CPU performance (driving clock speeds and straight-line instruction throughput). He suggests that this puts us at a fundamental turning point in software development, where increased reliance on concurrency is expected.
- In a November 2004 interview with HPCWIRE, Dr. Burton Smith, Cray's Chief Scientist said: "...the uniprocessor has pretty well run out of steam. Parallelism to date has been a nice strategy for HPC users and an afterthought for microprocessor vendors. Now, it is becoming a matter of business survival for all processor vendors. Parallelism is going to be taken more seriously, starting with the idea of exploiting multi-threading and multiple cores on a single problem. This is a major change. Imagine if Microsoft wanted to write Office in a parallel language. What would that language be, and what would be the architecture to support it? We don't have good answers to these questions yet." XMT provides answers to these questions about language and architecture.

A more direct introduction to XMT follows. Research over the last two decades by academic algorithm designers has produced a huge knowledge-base of parallel algorithmic methods, which is arguably second in its magnitude only to serial algorithms. The model of parallel computation used for developing this knowledge-base is called PRAM (for parallel-random-access machine, or model). Non-expert obeservers will probably be most impressed with the fact that superiority of the PRAM model, relative to other parallel computing models, was established through a fierce competition among quite a few schools of thought. This unofficial competition has been very well funded by government and industry and involved participation by thousands of the most talented computer science and engineering PhDs over several decades. In other words, arbitration through Darwinistic-type natural selection among different parallel algorithmic models is already behind us: the PRAM model won by a wide margin. However, this left open a fundamental question:

The non-expert observer may want to take note that our positive answer came as a big surprise to many, including parallel computing pundits, who have long argued that PRAM algorithmics is "unrealistic". The surprise may even be bigger since XMT relies on many of the great ideas that have been developed by parallel computing researchers. For this reason, we feel that, first and foremost, XMT should be viewed as a vision of triumph for parallel computing and for those who have labored for many years towards the goal of reducing single task completion time by way of parallelism.

To name a few: Schwartz's Ultracomputer vision of the PRAM as a theoretical yardstick, the NYU-Ultracomputer with its Fetch-and-Add primitive, the Shiloach-Vishkin PRAM work-time algorithmic framework (harnessing Brent's appearingly too-abstract theorem), NESL's system-supported nesting of parallelism, single-program multiple-data (SPMD) programming languages, Data-Flow's Fork-Join, the SB-PRAM multi-processors, the HEP-Tera thread switching, the IBM-360 out-of-order execution, Cray's compiler-based vector parallelism, the U. Wisconsin-Multiscalar multiple program counter (PC) architecture, and the U. Washington Simultaneous Multi-Threading (SMT) technique for mixing instructions from several PCs for a more effective use of functional units. The broad XMT framework seeks to build on these ideas, and attempts to make use of the ``Billion-transistor-chip'' technological opportunity towards an eventual resolution of the two questions above. The above names and efforts are meant to demonstrate that XMT should be viewed as an enhancement of past great ideas in parallel computing, and to build on and promote this past work. XMT aims to prove that parallel computing, as a field, was right! It should be noted that perhaps the most common misunderstanding about XMT is that ``it is only an architecture''. In fact, XMT is a fine-grained parallel computational framework covering the spectrum from algorithms through architecture to implementation. Architecture is a crucial, though relatively small, component: the main implementation related innovation in XMT was through the incorporation of low-overhead hardware and software mechanisms (for more effective fine-grained parallelism).
DISCLAIMER: The books by Almasi and Gottlieb, by Culler, Singh and Gupta, and by JaJa provide a proper scholarly treatment of parallel computing and describe the evolution of ideas in the field. The above names are only for demonstration purpose.

Expanding on the first question. 1-Billion transistor chips, up by a factor of more than 30,000 (from 30,000 transistors) sometimes in the 1980s, are becoming a reality. What should we do with all this hardware? Many will agree that the standard computing framework (e.g., the one which is based on Intel's X86 instruction set architecture platform) does not provide a good answer and there is need to considerably update it. The XMT proposal aspires to provide an answer. (For a pointer to other schools-of-thought which address the same motivating question see: Computer, Volume 30, Number 9, September 1997.) A first priority in the design of XMT has been to harness the computing power of such chips for the purpose of reducing the completion time of single computation tasks.

Expanding on the second question. Explicit Multi-Threading (XMT) is a new conceptual framework for general-purpose computing. The (virtual) thread structure of PRAM-like algorithms is very dynamic: the number of threads that need to be generated changes frequently, new threads are generated and terminated frequently, and often threads are relatively short. Perhaps the most distinguishing feature about the XMT framework is that it envisions an extension to a standard instruction set which aspires to efficiently implement PRAM-like algorithms; XMT does so using explicit multi-threaded instruction-level parallelism (ILP). In XMT, we tried to minimize changes to current practice: new elements are added to computer languages, architecture and implementation, only where needed.

Newsletter story