Learn More
In this paper we used detailed simulations of application kernels to show that memory contention can substantially degrade the performance of SPMD computations on large-scale shared-memory multiprocessors. We showed that under row-major allocation, memory contention is due to synchronized access to entire rows of a matrix, rather than simultaneous accesses(More)
(3) where k is the network radix, n is the dimension, and k d is the average distance a message travels in a dimension. For a torus with channels in both directions, k d = k=4. B Description of Tmul-T Trace File Format The trace le is a sequence of 6-byte packets. The rst byte is the processor number and the second byte is an opcode to be described below.(More)
  • 1