Note: this article is part of the inside SC09 series of features we are running in the weeks before SC09 this November. In this series of articles we are talking with some of our community’s leading thinkers about what they are doing today, where they are headed tomorrow, and the outlook for HPC. It is our hope that this series will help inform your own view of HPC’s future, and make your time at SC09 more productive and rewarding.
Stan Ahalt is still excited about what he does, a fact that still surprises (and inspires) me every time we talk. This energy probably explains a good bit of the success he has enjoyed in his career. Now the head of the Renaissance Computing Institute (RENCI) in North Carolina, Ahalt was until recently executive director of the Ohio Supercomputer Center, and before that a professor in the department of electrical and computer engineering at The Ohio State University for 22 years. While at OSC he started the Blue Collar Computing initiative, a national program aimed at bringing HPC to nontraditional users in business and industry, and OSCnet, a high-speed research network for K-12 schools, higher education and economic development.
On the national stage he is the chair of the Coalition for Academic Scientific Computing, co-chair of the Ohio Broadband Council, extramural member of the National Cancer Institute’s Advanced Biomedical Computing Center’s Oversight Committee, and member of the Council on Competitiveness High Performance Computing Advisory Committee.
Looking ahead at RENCI
When I spoke to him in late September, he had just moved in to his offices at RENCI. RENCI is facing some tough times these days: a 35% budget cut was announced just days before he took the job, a result of a decline in state revenues resulting from the tough economic climate.
The $3.8M cut (the media was inaccurately reporting $11M for a while) forced the elimination of 23 jobs, but Ahalt was aware of the possibility of cuts before he accepted the position, and the challenge hasn’t slowed him down. “RENCI is an assemblage of remarkable resources,” he says. “And we have had consistent state support even though the budget was cut. This was just a reaction to the current financial reality; lots of budgets got cut.” He describes RENCI as having incredible opportunities for the future, all driven by “the incredibly talented people at RENCI.” But his resources are now less than they once were. His reaction? Focus.
Ahalt says that RENCI is going to focus on two key areas of its mission. The first, environmental planning and emergency preparedness, is an area that Ahalt sees as already world class at the Institute. This work includes topics as diverse as storm surge analysis, fire prediction, flood plain studies, and many others. The second area, health care and medicine, is a growth opportunity for RENCI, and capitalizes upon the wealth of research being done by its neighbors, including Duke and UNC Chapel Hill.
Our biggest challenge? Software.
Our conversation about the key challenges and opportunities facing our community was broad. Our leading challenge? Software. “Since 1999,” he says, “all of the PITAC and PCAST reports have emphasized software investment needs. But during that time software funding has steadily declined relative to hardware funding.” Where are the main opportunities here? He points out that ISV codes still don’t scale well past a small number of processors, and we have substantial needs in middleware and “ease of use” systems like portals. According to him the consensus at the recent CASC meeting was that large-scale application codes will probably need domain-specific languages. “DOD and DOE are pretty effective today,” he says, “but small and medium-sized companies need help. We need to push HPC into new places.”
What about software needs specific to exascale systems? Interestingly, he shares the circumspect attitude Justin Rattner had when I talked to him last week about this same issue. In particular, he is impressed at the current performance of the non-application suite of tools that power today’s petascale systems. “We are having problems,” says Ahalt, “but not nearly as many as we thought we’d have.”
Workforce of tomorrow
Another key area of concern for Ahalt is HPC workforce development. But not at the high end: “We are doing OK on staff to take care of the very large scale HPC installations,” he says. “But we need to focus on building a large, well-trained community of users for the ‘everyday’ HPC. This is a national competitive issue. We lead in cycles, but we don’t have the minds to use them.”
This has been a topic of conversation for years now, and I was curious to know what approach he thought would work, and why we haven’t already made more progress. My computational engineering degree — still a new discipline in the early 90s — was supposed to have addressed some of these problems, but it hasn’t. “We probably need a national level effort,” he says. “We know how to fund hardware, but we haven’t really tried hard enough yet to pull people into computational science.”
Data — lots and lots of data
A third area in which he sees both challenges and opportunities for our community is in finding, storing, labeling, and exploring the hundreds of exabytes of data modern society produces each year. My own recent experience as a data center manager is that much of the science and engineering community isn’t “managing” their data at all. They simply write it all, every byte they generate, to vast archival storage systems. But without the tools to label, search, and dig into their growing archives, the reality is that most of what they write is never accessed again.
As Ahalt sees it, this free ride has to come to an end. “I think we are fooled by our personal experiences with our laptops and desktop computers that says that we can always buy another terabyte to store the extra data,” he says. “Computer cycles are a mechanism to get more actionable information than we started with,” and engineers and scientists who create massive amounts of data need to seriously examine when the whole dataset needs to be saved, when only parts of it should be stored, and when it is simply more cost effective to recalculate if data are needed again in the future.
He does see some hope that other factors in the computing community may drive the changes in user behavior needed to begin managing the data deluge. “Data movement has always been a problem, but as some classes of users shift to cloud computing they will be dealing directly with data cost and retention issues. This could be a good thing,” finally exposing the implicit costs of large scale data storage and management.
Hosted computing is most commonly lumped in with the term “cloud computing” these days, but the idea is old. Some types of capacity-oriented HPC jobs will run quite satisfactorily on clouds that are designed specifically for this kind of work (like Penguin’s POD offering; there is already mounting evidence of the inappropriateness of general and highly-virtualized cloud solutions like Amazon’s EC2 for HPC work).
In terms of this capacity work Ahalt feels that “a dramatic percentage of this type of HPC will fall naturally into a cloud computing model.” This is not to say that everything will migrate to the cloud: “We will still need [NSF] Track 1 and Track 2 machines. But what about a ‘Track 3?’ There isn’t one, but we still need more workhorse machines.”
The greater good
As Stan Ahalt gets ready to lead RENCI into the next phase of its life, he is also preparing to help lead the HPC community through these, and other, challenges. I asked him to describe his role at RENCI in this context. For the state of North Carolina, Ahalt is the go-to guy for HPC. But as RENCI’s new director he is focused on helping the Institute meet its mission of making a difference for society through technology. That will mean wrangling desktops, GPUs, clusters, supercomputers — and many other technologies — into the service of the greater good.