is leading to speculation about new models of computing, the technology’s overhead remains a core challenge. Recent advances in hardware and software have been removing some of the performance concerns associated with virtualization, but the goal is to eliminate the performance gap altogether. “We are not there yet, but what you’re going to see is enhancements in processors and other technologies to make the performance gap go away,” says Leendert van Doorn, who is a senior fellow at AMD and responsible for AMD’s virtualization technology, including the AMD virtualization extensions in the company’s latest quad-core Opteron processor, which are designed to reduce the performance overhead of software-based virtualization. “The big problem with virtualization right now is performance guarantees,” he says. “If you have a database transaction requirement of a few milliseconds, it is very difficult to provide that guarantee in a virtualized environment.”
Still, van Doorn says he is confident that this overhead will be reduced in the coming years with better hardware and software support for virtualization. Currently, overhead in virtual-
ized environments varies from a few percent to upward of 20%, a figure that van Doorn says depends on several factors, including how the hypervisor is implemented and whether the operating system running atop the hypervisor is aware that it is being virtualized. “The Holy Grail is to get near-native performance,” he says. “We are getting closer to that goal.”
In addition to the performance issue, there remains the issue of man-ageability in the data center and elsewhere. “For the next generation, every big software company is working on comprehensive management tools,” says van Doorn. The goal is to deal with a massive number of virtual machines
and effectively make global optimization decisions for thousands of virtual systems running in data centers or in the hands of a large work force. Sophisticated management tools will be essential in the future imagined by virtualization’s proponents, who predict that industry is moving toward a world in which the technology is ubiquitous, and where all new machines will have virtualization capabilities embedded in firmware.
Certainly, says Citrix’s Pratt, all servers, desktops, laptops, smartphones, routers, storage arrays, and anything else running software that must be isolated from other applications will be virtualized. The result? “The main noticeable thing will be more trustworthy computing,” says Pratt. Echoing this sentiment, Herrod predicts that users won’t think about virtualization as a different form of computing. “It will seamlessly fit into our notion of computing,” he says, “enabling a much simpler and more productive experience for all of us.”
Based in los angeles, Kirk L. Kroeker is a freelance editor and writer specializing in science and technology. steven hand, citrix, and carl Waldspurger, Vmware, assisted in the development of this article.
the world of computer science recently lost two esteemed members: oliver G. Selfridge, who died at 82, and ingo Wegener, 57.
Selfridge, whose career included positions at Mit,
BBn, and Gte
laboratories, is
widely regarded as
a leading pioneer
in the field of
artificial
intelligence and
the father of machine perception.
“in prescient research in the
1950s,” says eric horvitz,
president of the american
association of artificial
intelligence, “he introduced and
tackled key problems that are
now well known to machine
learning researchers, including
the challenges of search and
optimization over large
parameter spaces, feature
definition and selection, dependencies among variables, and unsupervised learning— learning without explicit access to signals about success versus failure.”
in 1956, Selfridge, with four colleagues, organized a conference at Dartmouth College that led to the creation of the field of artificial intelligence. and his 1958 paper, “pandemonium: a paradigm for learning,” is a classic ai treatise that essentially provides a blueprint for machine learning research.
“the pandemonium work introduced a distributed model for pattern recognition, where a community of interacting ‘demons’ or agents with different competencies and functions perform different subtasks that are then combined into final answers or behaviors,” horvitz notes. “Rather than
metaheuristics, and his
conviction that optimization
algorithms based on
metaheuristics, like evolutionary
algorithms and simulated
annealing, should be studied
with the methods from
the theory of
efficient
algorithms and
complexity theory.
Wegener’s new,
theoretical
approach
produced a profound understanding of the limitations of such metaheuristics.
Wegener was appointed a member of the German Council of Science and humanities, the leading scientific advisory committee to the German government, in 2004, and won the Konrad-Zuse-Medal, Germany’s most prestigious computer science award, in 2006.
left: PhotograPh courtesy of caroline selfridge, right PhotograPh courtesy of informationsdienst Wissenschaft
being handcrafted ahead of time and fixed, the agents and their networks of communication could evolve with experience.
“For decades, oliver communicated an exciting vision where computers would one day learn to infer human intentions and act to assist people without the need for detailed expression of problems,” says horvitz. “Such a vision has evolved to be central in research on human-computer interaction.”
ingo Wegener, a professor of computer science at the technical University of Dortmund, is well known for his groundbreaking work in complexity theory. he wrote a pair of important monographs,
The Complexity of Boolean Functions (1987) and Branching Programs and Binary Decision Diagrams
(2000). in the early 1990s, he worked in the formal analysis of
References:
Archives