This week Yahoo, HP and Intel announced their contribution to what is becoming an active competition to develop the infrastructure for next generation computational services. The announcement of the Cloud Computing Test Bed is broader — in scope and scale — than the previous IBM-Google announcement establishing the Cluster Exploratory (CluE), and all this competition is definitely a good thing for the academics who will have several years of work on these resources. If the research efforts bear fruit, they may also help shape the future of HPC as well.
The idea behind these research efforts is that the emerging model of large, generally available compute resources, and the increasing demand for “Internet-scale” applications, will not be satisfied by existing approaches to building infrastructure and applications. Fundamental changes — as yet unknown, hence the need for research — will be necessary in computer science and software engineering from the operating system up to the application.
On Tuesday Yahoo, HP and Intel announced the creation of a new open test bed for “the advancement of cloud computing research and education.” The physical dimensions of the test bed span six centers on three continents, with hardware at the Infocomm Development Authority of Singapore (IDA), the University of Illinois at Urbana-Champaign, the Steinbuch Centre for Computing of the Karlsruhe Institute of Technology, HP Labs, Intel Research and Yahoo. Each center will have HP/Intel hardware — the companies only refer to “infrastructure,” so it’s unclear whether there will be one large cluster at each site or a farm of machines — ranging from 1,000 to 4,000 cores that will be used to support cloud software research.
Representatives from the three companies held a conference call on Tuesday to talk about the test bed. On the call were Prith Banerjee, senior vice president or research at HP and director of HP Labs; Prabhakar Raghavan, head of Yahoo Research; and Andrew Chien, vice president of the corporate technology group at Intel and director of Intel Research. During the call, Banerjee said that the main goal of the test bed is to remove the financial and logistical barriers that might otherwise prevent people from developing effective cloud computing applications.
Raghavan said in a prepared statement that “With this test bed, not only can researchers test applications at Internet scale, they will also have access to the underlying computing systems to advance understanding of how systems software and hardware function in a cloud environment.”
Google and IBM announced a joint venture to provide “hardware, software and services to augment university curricula and expand research horizons” into ways of building applications that can live happily (and provide effective service to users) in a cloud environment. This effort was announced in October of last year with a pledge of 1,600 processors in Google and IBM datacenters. Google and IBM announced at the time that they were partnering with a small number of universities in the U.S., including the University of Washington, Carnegie Mellon, Stanford, MIT and others. In February of this year, the NSF joined the effort in a role that manages the stream of supplicants seeking access to the resources for research, something one would expect the NSF to be very good at.
When asked why the partners in this new test bed didn’t just join up with the Google/IBM effort, Chien said that while the two efforts are complimentary, they are different in important ways, “What we’re trying to do is support research at a variety of levels of the software stack, not only at the application layer, but also down in the system software, in the manageability, and eventually exploiting some of the novel hardware platform features…. my understanding is that the Google/IBM partnership is focused primarily at the application level.” That’s my understanding, too, by the way.
The companies claim that parts of the datacenter infrastructure are up and running now, with more coming online throughout the year. During the discussion about the effort, Raghavan of Yahoo referred several times to the M45 center being “already up and running,” as if that center is to become part of this effort. The precise relationship of that datacenter to this effort was not clearly articulated, however, and we’ll have to wait for further details from the company.
Something else that wasn’t made clear in this announcement is how resources are to be allocated to institutions and individuals that want to build on the platform. It was several months before Google and IBM announced their relationship with NSF that contributes to the “researcher management” aspect of that project, so we may have to wait a while to find out how that’s going to be handled. Chien, in response to a question about whether others would be allowed to join, indicated that this week’s announcement is the “leadership step” and that “we believe that the larger infrastructure we can get together, the more valuable it will be to the research community, so we’re open to expanding this and having other folks join.”
Yahoo of course has a history of this sort of thing, having announced a collaboration with the Indian Tata Group for large-scale cloud computing research to the tune of 14,000 cores, and the M45 datacenter, a 4,000 core machine dedicated to “large-scale systems software research.” Intel, too, has sponsored efforts that are related to the Cloud Computing Test Bed, including the recently announced Universal Parallel Computing Research Centers in partnership with Microsoft.
Of course, not everyone is uniformly enthusiastic about either the Google or Yahoo announcements. Comments on some of the well known Internet news sites (including the Bits blog at the NY Times) point to what they observe is an accelerating pace of change in the acronyms and buzzwords used to describe old problems that never actually get resolved before someone rebrands them and everyone heads off in a “new” direction.
So, what might all this mean for HPC? Again, I turn to the “million monkeys coding” corollary of the better known “infinite monkeys typing” theorem. Sure, software programmers aren’t monkeys, but the more of them we have working on large-scale software problems, the better our chances that we’ll develop a new set of approaches to developing high performance applications.