Parallel Computer Architecture

868 CHAPTER 11 Latency Tolerance

important for a system that allows write

—>

read reordering, particularly when the L

2

cache is lockup-free, but less important for the system that allows write

—»

write

reordering as well. The reason is that in the former case writes retire from the buffer

only after completion, so it is quite likely that when a read misses in the first-level

cache it will find a write still in the write buffer, and stalling the read until the write

buffer empties will hurt performance. In the latter case, writes retire much faster, so

the likelihood of finding a write in the buffer to bypass is smaller. For the same rea-

son ...