Startup Aims to Bring Parallel Applications to the Masses
There are a number of young companies at SC11 this week, debuting novel technologies. One of them, Advanced Cluster Systems (ACS), recently launched its first software product, with the rather bold name of Supercomputing Engine Technology (SET). It promises one of the Holy Grails of HPC: to turn sequential applications into parallel ones.
HPCwire got the opportunity to ask ACS founder and CEO Zvi Tannenbaum and, ACS CTO Dean Dauger about the product and the underlying technology.
HPCwire: How did Advanced Cluster Systems start out?
Zvi Tannenbaum: ACS was established in November 2004. At that time, ACS’s mission was to expedite scientific processes for a particular finance company utilizing Wolfram Research gridMathematica. While working on solutions, I realized how hard it is to do (Parallel Mathematica is grid-based, not supercomputing-based), and the great expense of software licenses, especially considering that the technology did not take advantage of all available resources (I.e., Mathematica utilized only half the cores available). I followed up with Wolfram, who explained that their software was not built to take advantage of multicore processors.
In late 2006, I contacted Dr. Dean Dauger, a world-renowned supercomputing expert who developed the easy-to-use patented Pooch clustering software, and explained to him my idea to attach Mathematica kernels to his Pooch clustering solution to give supercomputing-like all-to-all connectivity to Mathematica kernels. Dr. Dauger finished this project in one month, practically turning Mathematica into a supercomputing application, without changing a single line of its source code — Mathematica is a proprietary software.
Once we finished the project, Dr. Dauger and I realized that we came upon something much bigger than just parallelizing Mathematica. We took a modular sequential code like Mathematica and provided it with supercomputer-like parallelism. We then continued our development and created SET.
ACS recently acquired the patented Pooch clustering technology from Dr. Dauger, who has joined ACS as an owner. ACS today holds the rights to three patents: two regarding easy-to-use clustering — the Pooch technologies– and the third regarding the recently granted SET patent. And ACS has other patents still pending, so more on the way!
HPCwire: Dean, can you tell us what the Supercomputing Engine Technology does and how it works?
Dean Dauger: SET applies the parallel computing paradigm of distributed-memory MPI, proven over the last twenty years to achieve efficient parallelism from multicore to clusters to clouds and supercomputers. However it has three defining differences from MPI.
The first is that it provides a support architecture and framework that covers common parallel computing patterns. Beyond simply message-passing patterns, SET “owns” the data to be manipulated across the parallel computer so that SET can organize and rearrange the data as needed for the parallel computing pattern. SET supports parallel data structures, such as partitioning with guard cells and element management, and parallel execution patterns, such as divide-and-conquer array generation, common to many parallel codes, including MPI codes.
Because that support has never made it into MPI itself, every writer for MPI has had to rewrite the same parallel data structure and execution again and again. SET makes parallel code writing easier by writing that once and not requiring users to debug that part of the code.
The second defining difference from MPI is that it has the application organized into a “Front End” and “Back End”, with distinct purposes. The Front End is the “captain” of the application, directing the entire application and making global decisions, much like the main function or main loop of a code. The Back End does the grunt work, the raw and low-level calculations.
SET is the bridge between the Front End and the Back End, but that division allows SET to organize the work performed by the many Back End codes as appropriate for parallelism. In particular, SET runs many Back End codes simultaneously, allowing the writer of the Back End code, because by definition it simply does its own chunk work on its chunk of data, to not have to think about parallelism.
The third is that the result is a parallel computing approach that is much easier to use for application developers. As much as possible, the details of parallel computing is handled by SET, whether it be data or execution management across the cluster. The application-specific pieces are in the Front End, which defines the high-level execution of the parallel application, and the Back End, where the low-level calculations are actually performed.
I personally enjoy MPI, but I’ve encountered many with parallel computing needs that see MPI as too much like assembly language. We designed SET with the scope necessary to cover parallel computing details while enabling the application writer to think sequentially as much as possible.
HPCwire: What types of dependencies does SET have on the underlying platform — OS, software stack, hardware, and so on?
Dauger: Fundamentally, SET needs a parallel system with some equivalent of MPI_Irecv, MPI_Isend, and MPI_Test, plus the usual metrics of the system (rank and size). This makes it possible to port SET to shared-memory as well as standard MPI systems.
At present implementation of SET runs on all the major Unix-compatible platforms. We’ve run it on OS X and 64-bit Linux clusters as well as larger systems like SGI. As ACS’s resources allows, we will expand SET to other OS’s.
HPCwire: How would a sequential program need to be modified so that it could tap into the SET technology?
Dauger: The application would be organized into a “Front End” and “Back End”: The Front End is the “captain” of the application, directing the entire application and making global decisions, like the main loop I mentioned before. The Front End is also where the user-interface, if any, resides.
The Back End does the grunt work, the raw and low-level calculations. Any modern modular code should be able to be factored relatively easily this way, as it is an excellent and well-accepted approach for reusing code between projects.
HPCwire: How long would this typically take?
Dauger: Factoring the application into the Front End and Back End should be straightforward for a modular or other modern, well-organized application. After that one adds “glue code” between SET and the Front End and SET and the Back End, which typically consists of wrapper calls or minor replacements. Then there’s testing and optimization. Most projects using conventional approaches allows a year to accomplish this. With SET this can take under a month.
HPCwire: How well does the technology scale in the multicore, multiprocessor, and multi-server dimensions?
Dauger: Since the underlying paradigm is that of distributed-memory MPI, it scales almost as well as distributed-memory MPI on all parallel computing implementations. Where SET might do poorly is also where other parallel approaches do poorly, such when communication time is far greater than the computation size. The purpose of SET is to make it much easier for the software writer to quickly produce an application that can achieve scale.
HPCwire: Compared to a hand-coded MPI application, how well does SET perform?
Dauger: The SET approach has produced codes that scale almost as well as traditional hand-coded MPI applications. In some cases the results are indistinguishable from what is accomplished via MPI.
HPCwire: Has the SET technology been applied to any real-world codes?
Dauger: The first major proof-of-concept is SET’s application to Wolfram Research’s Mathematica. Mathematica is a very large application, millions of lines of code. The usual approaches would have taken a year, probably longer. Applying our SET to Mathematica took only one man-month, and yet it is able to scale far better than any other solution using Mathematica. That is now a product named Supercomputing Engine for Mathematica. Notably, because Mathematica is well modularized, we didn’t even need to look at Mathematica’s source code.
HPCwire: Zvi, what’s the company’s business model?
Tannenbaum: The primary sales strategy for Advanced Cluster Systems is to execute on a reseller channel program to leverage ACS by creating independent contractor relationships with value-added resellers — resellers or VARs — and solution providers. This will enable ACS to deploy a sales team with varied industry expertise, existing relationships with prospective customers and worldwide sales coverage without the fixed expense of hiring a direct sales force beyond a Director of sales/reseller channel manager.
HPCwire: What’s the next step for Advanced Cluster Systems?
Tannenbaum: ACS is a very small company and the next step is to cultivate its technology while remaining focused on our plans and avoiding pitfalls that can stop a tech-company’s growth in its tracks. In addition to continued development and executing plans to implement our Reseller Channel Program, we are working with major hardware companies to get SET exposure. We are currently working with an American VAR to distribute an upcoming SET-based enhancement to a scientific software product, and we are talking with cloud service providers.
We also have established our presence in Europe with Daresbury Labs and a British solution provider on site. We realize that a good reseller channel program must have a good marketing program to support it, and we are pursuing that course as well. We are also revising our business plan to prepare for external funding to achieve our goals and continue controlled and consistent growth.