The Cell is a very powerful computing platform that originated from a processor that was developed for wide application in a mass market consumer device, namely the recently introduced Playstation 3. The introduction of this new platform may create a new economy of scale on the high performance computing (HPC) market, with the potential of the relative inexpensive but yet very powerful Cell replacing or complementing more traditional existing HPC devices. The goal of this innovative project is to study to what extent the Cell platform is applicable for high performance scientific computing.
It is proposed to purchase as many Cell units (pizzas, blades, or actual playstations) as possible given the budgetary limitations, to form a reasonably sized cluster. This cluster of Cell nodes would be the first such cluster in Switzerland. In the future, a cluster of this sort could complement the park of HPC machines in Switzerland (installed at the CSCS, ETHZ, EPFL, and other institutions). It would be used to run those applications that are particularly suited for Cells, due to their specific architectural properties, thereby in an efficient and cost-effective manner liberating resources on other machines.
To realize this vision it is imperative to study in a timely and pro-active fashion through bench-mark testing and extensive production runs with real-life HPC applications, which of those applications can optimally profit from the powerful Cell architecture. Furthermore, the installation of a first moderately sized cluster of Cell nodes would be helpful to answer several questions before a bigger machine would be acquired:
1. Which of the present applications can fully take advantage of the power and architecture of the Cell?
2. How could the Cell architecture be modified to improve the Cell's performance for memory bound applications (such as computational fluid dynamics, plasma physics, sieving, and sparse matrix calculations)?
3. How adequate is the basic software (compilers, libraries, etc.) on the first generation Cells, and what additional software would be required or needed to be developed for the present applications to guarantee high performance on Cells?
This project involves various disciplines (mathematics, computer science, and physics) and continues the fruitful collaboration between the Laboratory for Cryptologic Algorithms (LACAL), the Laboratoire d'ingénierie numérique (LIN), and the Centre de Recherches en Physique des Plasmas (CRPP), all at EPFL.
|