Mainframe Services from Gigabit-Networked Workstations

Until recently, large mainframes and super-computers were considered essential for powerful scientific batch computing services requiring intensive tape usage, large well-managed disk storage systems, high throughput and maximum reliability. However, this situation has changed dramatically over recent years with the appearance of RISC-based workstations with performance characteristics, at least for scalar computations, comparable with the fastest mainframes but with an order of magnitude better price/performance. At the same time, competitively priced workstation-class disk and tape systems with adequate performance and reliability have become available. Combined with newly-developed LANs and Gigabit networking solutions, it is now possible to provide scalable and integrated mainframe-class services on workstation platforms with the UNIX operating system.

Previous papers have summarized CERN's work over the past two years in developing and introducing such services on a large scale. The latest system is called SHIFT or Scalable Heterogeneous Integrated FaciliTy. The SHIFT facility performs a wide range of scientific data processing tasks including many with high I/O requirements and is comparable in CPU capacity to the CERN computer center. Similar systems are now being built within the budgets of smaller institutes which previously had to depend on remote university or national computing centers.

The present paper gives a short review of the SHIFT project's goals and architectural principles, and a detailed account of the networking and software design and implementation problems that were encountered and solved.