After having achieved some level of success, most startup companies usually end up replacing the founding CEO with a more seasoned business executive. At Interactive Supercomputing Inc. (ISC), this transition took place on January 10th, when they swapped founder and CEO Pete Peterson with Bill Blake. Peterson will remain with the company as the executive chairman of the board.
Bill Blake has been involved with high performance computing systems for more that 20 years. Before coming to ISC, he was a senior vice president of product development for Netezza, where he lead a team that developed a terascale database appliance for real-time business intelligence, employing a parallelized version of SQL. Prior to that, Blake was vice president of high performance technical computing at Compaq, where he led development and marketing efforts of HPTC solutions, including the company's AlphaServer SC supercomputers.
The fundamental reason for this executive shuffling is to move the organization from startup company to one that is more actively pursuing growth and partnership opportunities. The company currently has over 30 customers for their Star-P software product. Star-P uses a client-server model to add parallelization to very high level language environments on the desktop (initially MATLAB) to enable applications to run on HPC servers. At this point, both SGI and HP servers are supported by the Star-P platform.
Beyond support for SGI and HP hardware, ISC would also like to partner with all the other tier one system OEMs. Since Blake is a member of the board of directors for Cray, it would safe to assume their are discussions going on there as well.
The most important extension to the Star-P platform will be the delivery of the Python client later in 2007. The company sees a lot of interest in the community for a Python front-end for technical computing. They believe adding parallelization to this language will broaden their client base substantially. Over time, Blake says they would like to support other very high-level languages such as Mathematica and R (a language for statistical computing).
Beyond 2007, there are some other ideas brewing. Since Star-P contains software hooks to connect external parallel libraries to user code, the company is looking into expanding this facility to integrate FPGA acceleration software functionality. The lack of maturity in software support for FPGAs or even more traditional co-processor accelerators is something that Blake thinks could be a real opportunity for ISC.
“This is a very interesting area now,” observes Blake. “In the next year or two, there are going to be some very fascinating system configurations, that will frankly require some sort of a reasonable software platform to allow people to move from traditional SMP and cluster environments into these accelerated environments.”
Some accelerator support may be done in conjunction with OEM partnerships. For example, HP's use of ClearSpeed boards in some of their deployments may be an opportunity to add support for this hardware.
While Star-P early adopters range from the Air Force Research Labs to Australia's CSIRO, they have also found a sweet spot in medical image processing. In this environment, computation is typically done on the desktop by neuroscientists using MATLAB codes applied to MRI scans. But as more advanced MRI machines are used to increase resolution and speed, computing requirements of many neuroimaging applications are outstripping desktop workstations. Since languages like Fortran and C with MPI are alien to this particular domain, Star-P provides a relatively easy way to get to HPC — assuming the lab can get access to some SGI or HP hardware.
ISC is also attracting the attention of Wall Street. Blake says they have one financial firm who has already bought the platform, although he could not disclose the company's name. Like other data- and compute-intensive workloads, financial applications are increasingly looking to HPC to get better performance and real-time turnaround. Derivative pricing applications, in particular, are being pushed beyond the limits of conventional spreadsheets and finance application tools.
The individuals who develop these application are quantitative analysts or “quants” for short. They typically have mathematical backgrounds as well as some computer science training. According to Blake, there is a fairly significant migration between the national labs and the high-end investment firms for people like this, especially as these financial organizations attempt to leverage HPC.
“The quantitative analysis that goes on there is really quite rigorous,” says Blake. “As they try to look at more financial instruments across multiple accounts, all of a sudden you're into serious high performance computing.”
Unlike medical imaging, financial applications use a fairly wide mix of language environments. C++, MATLAB and Excel are probably the most widely used, but parallel programming via MPI or OpenMP is not yet widespread. ISC is hoping that the Star-P/MATLAB combo will look increasing attractive to Wall Street as they attempt to leverage high performance computing for their financial instruments.
While most of ISC's early customers are in government/academia or the health/bio sector, over time they expect to participate in all of the traditional high performance computing vertical markets. But for now they will be content to consolidate their reputation in their established markets.
“Doing an adequate job in 20 places isn't as interesting as being extremely useful in the first two or three,” says Blake. “So we'll be extremely cautious on how we expand.”
Blake believes that ISC's focus on maximizing end-user productivity is in line with the changing culture of high performance computing. When he worked on the Alpha project at Compaq, MPI was used to achieve the highest level of scalability possible in order to maximize performance. In constrast, the goal at ISC is to abstract the front-end interface as much as possible and deliver the parallelism in the runtime. With the costs of software development so high relative to hardware, even performance-minded users seem willing to make some tradeoffs. Even if the application achieves somewhat lower levels of scalability or runtime performance, getting the answer to the problem with significantly less programming effort becomes very attractive. Blake is betting that the time is right.
“Very high-level languages are definitely a part of the productivity improvement that people are seeking, and I do support the notion of High Productivity Computing that Debra Goldfarb has talked about,” says Blake. “It's very real. It makes no sense to speed the computation up by orders of magnitude if you have to spend years writing those codes.”