clarke@utcsri.UUCP (Jim Clarke) (02/23/87)
(SF = Sandford Fleming Building, 10 King's College Road) (GB = Galbraith Building, 35 St. George Street) SUMMARY: Tuesday, March 3, 11 am, SF1101: Jerome Feldman (title TBA) Tuesday, March 3, 3 pm, GB 120: Jim Delgrande (title TBA) Tuesday, March 3, 4 pm, GB221: Josef Saad ``Sparse Matrix Computations on Parallel Machines" Thursday, March 5, 11 am, GB220: Michael Schwartz ``Naming in Large Heterogeneous Systems" Thursday, March 5, 3 pm, GB220: Gaston Gonnet ``The Use of Many-to-One Mappings for Algebraic Pattern Matching" ------------------- COLLOQUIUM, Tuesday, March 3, 11 am, SF1101 Professor Jerome Feldman University of Rochester Title: TO BE ANNOUNCED A.I. SEMINAR, Tuesday, March 3, 3 pm, GB 120 Professor Jim Delgrande Simon Fraser University Title: TO BE ANNOUNCED NUMERICAL ANALYSIS SEMINAR, Tuesday, March 3, 4 pm, GB221 Professor Josef Saad University of Illinois ``Sparse Matrix Computations on Parallel Machines" This talk will describe the difficulties encountered when solving large sparse matrix problems on parallel machines and propose a few tenta- tive solutions. We will start by presenting a brief overview of the dif- ferent techniques used (direct and iterative methods) as well as the data structures employed. These methods suffer from numerous bottlenecks when implemented in parallel or vector machines. We will focus on iterative methods which are usually much faster than direct methods for very large problems, although often not as reliable. Work remains to be done to improve the robustness of the standard preconditioning techniques by, for example, incorporating pivoting in the incomplete LU factorizations. On parallel machines, most of the standard preconditioners such as those based on incomplete LU factorization, are inherently sequential. A class of preconditioners that appears promising for general sparse matrices is that of polynomial preconditionings. We will explain this by comparing the per- formances of different ways of implementing matrix by vector multiplica- tions on the Alliant FX/8. Finally, a few preliminary tests with a code based on polynomial preconditioning on the Alliant FX/8 will be reported. SYSTEMS SEMINAR, Thursday, March 5, 11 am, GB220 Professor Michael Schwartz University of Washington ``Naming in Large Heterogeneous Systems" This talk will discuss an approach to naming in evolving systems that are composed of a heterogeneous collection of subsystems. Our goal is to reduce the cost of integrating new system types into an exist- ing environment. Providing a naming service in such an environment is difficult because each of the systems being integrated typically uses its own naming service. The obvious approach is to implement a new standard to which the existing subsystems must adhere. However, several problems with such standardization render it unsuitable for the environment we are addressing. We present a new approach based on integrating existing name spaces through a structure that separates global name space administration from knowledge of the semantics of naming in each of the assimilated subsystems. A major advantage of this approach is ease of integration: newly added system types can participate in the larger system without modification. At the same time, systems that use the global naming service can take advantage of the services provided by new systems without modification. In addition, the model has broader applicability than the applications typically associated with naming. It can, for example, form the basis for a file system that integrates a heterogeneous collection of existing file systems. A prototype implementation has been built in conjunction with the Heterogeneous Computer Systems project at the University of Washington, and is being used as a basis for developing several heterogeneous applications. Analysis of measurements indicates the applicability of caching at various levels of the system. THEORY SEMINAR, Thursday, March 5, 3 pm, GB220 Professor Gaston Gonnet University of Waterloo ``The Use of Many-to-One Mappings for Algebraic Pattern Matching" -- Jim Clarke -- Dept. of Computer Science, Univ. of Toronto, Canada M5S 1A4 (416) 978-4058 {allegra,cornell,decvax,linus,utzoo}!utcsri!clarke