Previous | Next --- Slide 18 of 52
Back to Lecture Thumbnails
paracon

Because of W-R reordering, the processor is able to hide the write latency by performing reads ahead of write commit.

jedi

The read latency is slightly higher in each of the relaxed W -> R versions, compared to the corresponding sequentially consistent executions. Why would that be?

sandeep6189

@jedi, I think both the Write and Read operations(note, both shown in yellow together) are both going in parallel for each W-R group. The small extra time would be some small communication/scheduling overhead. This is a good graph to demonstrate latency/memory hiding

o_o

By relaxing the memory consistency W->R ordering constraint, processors are able to almost completely hide the write latency by allowing reads during the write. Now that reading and writing happens at the same time, in the yellow bar, the read time takes slightly longer probably to cover for the writing that is happening in parallel.