Cluster Computing:

CONFERENCE CHAIRS DISCUSS THE IMPORTANCE OF CLUSTERING
by Tim Curns, Editor

HPCwire caught up with the general co-chairs of IEEE's Cluster 2004 Conference (http://grail.sdsc.edu/cluster2004/) Daniel Katz, of the Jet Propulsion Laboratory (JPL), and Henri Casanova, Assistant Research Scientist at SDSC, to ask some questions about the show and cluster computing in general.

HPCwire: To begin with, please give a brief overview of what you've seen or heard at Cluster 2004.....what's happened here in San Diego? What has impressed you? And how do these things compare to what you expected?

Henri Casanova: Being involved with the conference organization I did not have the chance to spend too much time in the technical sessions or even at the exhibit, so my view of the conference as an event is probably not representative of what a regular attendee would experience. I was personally impressed by the number of exhibitors that stepped in to sponsor the conference and showcase their technology. This of course helped the conference financially, but also contributed greatly to a more diverse and enjoyable experience. This has been above my expectations. This year we decided to have only 2 parallel technical sessions, which I believe worked much better than 3. It made the conference more focused, and the technical sessions livelier.

HPCwire: So Dan, what has impressed and or suprised you?

Dan Katz: The effort made by all the people to attend/present/exhibit/etc. It's all been more work than I was expecting.

HPCwire: Is this the first time for both of you chairing an event like this? What other experiences have helped you serve as General co-Chairs?

HC: It was my first time chairing an even like this!

DK: I've had a variety of experiences in the Cluster series, including being 1 of 3 Program Vice-Chairs, Deputy Program Chair, General Vice-Chair, Co- Program Chair, and now Co-General Chair. Cluster2001 was most helpful to me, because that meeting was organized by JPL, and it seems that the local people really see all of what's happening, where the remote people often only see their part of the conference.

HPCwire: Let's get into some cluster talk. Where do you see the line between cluster computing and Grid computing being drawn?

HC: The line can of course be defined from the traditional technical standpoints (in a machine room, with a single file system, over a switch, etc.). Today, the most common components of large-scale grids are clusters, and, rather than trying to partition the domains, lot of work is underway trying to understand what work can be done both on the cluster side and on the Grid side so that such platforms can be easily constructed, maintained, and used. So, there is a clear interface and overlap between these two worlds. Furthermore projects such as the Optiputer push the limit of the classical definition of a cluster as fast optical links allow for building clusters whose nodes are separated by long distances. This is where the line between Grid and cluster computing becomes very blurred, and it will be interesting to see what develops over the next couple of years.

DK: Clusters are not grids, though they can be elements in grids. On the other hand, clusters can be used in a manner similar to grids, though doing so seems to be a waste of resources. The difference is the networking - clusters are much more tightly bound together than grids. Also, clusters are intended to be a single system, while grids are intended to be used separately, but of course can be used together.

HPCwire: So how do you see SDSC and UCSD fitting into these areas?

HC: SDSC and UCSD are leaders in both areas and involved in projects such as TeraGrid and Optiputer, that both involve deep cluster computing and Grid computing issues, spanning the whole spectrum between the two, and contributing to defining where the field is going.

HPCwire: What exactly makes PC clustering easier to utilize for classical supercomputer users, as well as experimental data processing groups and bioinformatics users?

HC: We come from a world where users/programmers used to gain immense experience developing applications on particular supercomputers, and used to go through painful transitions each time a new generation of machines would be available, possibly from a different vendor, with a different architecture, with a different operating system, and different compilers. By contrast, there is a large effort today to standardize cluster computing (as exemplified in the Beowulf idea), a lot of community code is available that can run on such clusters, entire frameworks are available for cluster management (e.g., the ROCKS project), and Linux (or flavors of it) has emerged as a de-facto O/S. This provides a much more uniform development environment, which is affordable to a much larger community of users, and which comes with a large and common software and knowledge base. In fact, Dr. Ryutaro Himeno, form the Riken supercomputer center in Japan, who was invited as a keynote speaker at Cluster'04 in San Diego, presented a very compelling case of why they made the decision to move from vector supercomputers to clusters for their scientific users, how it made sense in terms of cost, of performance, and of ease-of-use for users. His presentation at the conference is only one of many examples of how clusters are a viable alternative to supercomputers for many classes of scientific users, even in settings in which entire generations of users have been trained on traditional supercomputers.

HPCwire: In addition to moves like Riken's, how do you think parallel processing and clustering has changed over the last decade or so? What issues are more prominent this year at Cluster 2004, than last year for instance?

HC: Cluster computing has become mainstream over the last decade, initiated by the Beowulf project, and we can see this today during the conference (not being able to attend Cluster'03 and Cluster'02, I cannot comment about what issues are more prominent this year). Of course, in terms of parallel processing, the main evolution has been Grid computing, and clusters are the most common components of a Grid platform.

DK: Obviously clusters are a new feature in parallel processing over the last 10 or so years, as they have now basically driven off many more powerful systems. The consequence of this is that only those applications that run well on clusters are being pursued, and much of work in new applications is stagnating.

The RAIT workshop (RDMA Applications, Implementations, and Technologies) was new to Cluster this year, and had about 40 attendees. This seems to be a new area of growth for the conference and for the world of clustering in general. Otherwise, the technical area of clusters seems fairly similar to last year.

I would like to see some discussion of how potential new languages and programming schemes might impact the world of clusters, but this didn't really happen in the technical papers this year. Perhaps next year in Boston...

HPCwire: How important is benchmarking, such as LINPACK for example, in the world of high-performance clustering? Do you think we should continue to focus on these litmus tests? Why or why not?

HC: Of course a benchmark such as LINPACK can only give a single and narrow view of the performance delivered by a machine (for instance, it does no I/O). However, it has value because it fosters competition and provides us with a view of the trends and evolutions of parallel computing platforms over the years. In fact, mining the data contained in the Top-500 lists over the last decade reveals a lot of interesting facts that confirm or disprove our assumptions about how technology evolves. So, in this sense, having something like the LINPACK benchmark and the Top-500 lists has a great value for the community. Now, I don't believe we, as a community, actually do "focus" on these tests. It is well understood what these tests are and what their limitations are. In fact, there are entire suites of benchmarks that, while not as visible as the LINPACK benchmark, provide much more complete views of a platform's performance and potential for real applications. Also, although performance modeling is always a very difficult question, there is a lot of expertise and good rules of thumbs in the community to assess how good a machine is for a given application, without taking things like the LINPACK benchmark into account necessarily. So, yes, we should keep such tests, and no, we shouldn't focus on them, which I don't believe we are doing as a community.

DK: Benchmarks are clearly important, but they are also clearly misused. The best benchmark is customer-specific, but it is unreasonable to expect vendors to run customer-specific benchmarks, except for very large procurements. I think we will continue to have benchmarks play an important role, as they serve an important function, but I hope to see the reliance on LINPACK reduced in favor of newer benchmarks such as the HPC Challenge benchmarks. These tend to represent a wider variety of potential applications, and the results of this suite should help customers determine what systems meet their needs.

HPCwire: Henri, before your work in the US, you worked at France's Ministry of Defense and as a graduate research assistant at the Institut de Recherche en Informatique de Toulouse (IRIT) in Toulouse, France. How has your work here differed? How prepared is the U.S. to handle the emerging importance of clustering and Grid computing?

HC: One of the differences I could see when I first started working in the US, and which is definitely relevant to the cluster computing community, is the fact that in the US there was at the time a greater synergy between academia and industry. In fact, I believe that the US is in an great position to handle the future of cluster and Grid computing precisely because of the tight connections and collaborations between academia, supercomputer centers, national labs, and industry. This is exemplified by projects like the TeraGrid for instance.

HPCwire: How about France's position?

HC: I cannot place a judgment on how France is positioned as I have not been closely involved with research programs there, although I collaborate with a few individual researchers.

HPCwire: What kept you guys up at night while organizing this event? What may continue to worry you in general regarding this field?

HC: Obviously, as a general chair, there are a lot of issues regarding logistics and budget, which are not specific to the IEEE Cluster conference but apply to all conferences. In terms of the Cluster conference specifically, I think that one important concern was that, in the face of the multitude of conferences in this and related field, Cluster would retain its identity as a conference truly focused on cluster computing, that would attract practitioners both from industry and academia, rather than aiming for a broader range of topics that would then make it undistinguishable from other conferences in this and related fields. In this context, one key issue was to attract a good number of industrial exhibitors who would demonstrate latest cluster computing technology. Luckily Cluster'04 has been particularly successful in attracting industrial sponsors and exhibitors, and is definitely the venue for Cluster Computing research and practice.

DK: Our financial situation was the most stressing, more than anything else, though this was based on not knowing the numbers of attendees before the event started, because the percentage of attendees who register on-site is so high. In general, Clustering will do fairly well in the future. My concern is for the applications that are not well served by Clustering.

HPCwire: Understandably. Is there anything else that either of you would like to add?

DK: If possible, I would like to publicly thank the Cluster2004 team, especially Greg Bruno (SDSC), our local arrangements chair, and Jim Ang (SNL) and Kurt Keville (MIT), our exhibits/sponsors co-chairs. The conference would not have been successful without their hard work.

Clearly, clustering is still as important as ever. And, with the emerging technology involved with Grid computing, we are sure to see even more advances in the area over the next several years...