DocumentCode :
2779376
Title :
Asymptotically tight bounds for computing with faulty arrays of processors
Author :
Kaklamanis, C. ; Karlin, A.R. ; Leighton, F.T. ; Milenkovic, V. ; Raghavan, P. ; Rao, S. ; Thomborson, C. ; Tsantilas, A.
Author_Institution :
Aiken Comput. Lab., Harvard Univ., Cambridge, MA, USA
fYear :
1990
fDate :
22-24 Oct 1990
Firstpage :
285
Abstract :
The computational power of 2-D and 3-D processor arrays that contain a potentially large number of faults is analyzed. Both a random and a worst-case fault model are considered, and it is proved that in either scenario low-dimensional arrays are surprisingly fault tolerant. It is also shown how to route, sort, and perform systolic algorithms for problems such as matrix multiplication in optimal time on faulty arrays. In many cases, the running time is the same as if there were no faults in the array (up to constant factors). On the negative side, it is shown that any constant congestion embedding of an n×n fault-free array on an n×n array with Θ( n2) random faults (or Θ(log n) worst-case faults) requires dilation Θ(log n). For 3-D arrays, knot theory is used to prove that the required dilation is Ω(√log n)
Keywords :
fault tolerant computing; systolic arrays; 2-D; 3-D; asymptotically tight bounds; congestion embedding; faulty arrays of processors; knot theory; low-dimensional arrays; random fault model; route; sort; systolic algorithms; worst-case fault model; Bridges; Computer science; Contracts; Fault tolerance; Laboratories; Mathematics; Military computing; Routing; Testing;
fLanguage :
English
Publisher :
ieee
Conference_Titel :
Foundations of Computer Science, 1990. Proceedings., 31st Annual Symposium on
Conference_Location :
St. Louis, MO
Print_ISBN :
0-8186-2082-X
Type :
conf
DOI :
10.1109/FSCS.1990.89547
Filename :
89547
Link To Document :
بازگشت