Parallel Computer Architecture

7.7 Clusters and Networks of Workstations 513

is due to the L

2

cache miss penalty and the remote address translation. The remote

write latency is essentially the same in the two machines. The prefetch cost is

improved by roughly a factor of two, obtaining a rate of one word read every 130 ns.

Each of the memory modules can service a read every 67 ns. Nonblocking writes

have essentially the same performance on the two machines. The block transfer

capability of the T3E is far superior to the T3D. A bandwidth of greater than 300

MB/s is obtained without the large start-up cost of the block transfer engine. The

bulk write bandwidth is greater tha ...