Relative performance of operations (aka "things that are slow")

In general, our Westmere X5650 processors are pretty fast. Some operations, however, are surprisingly slow. This page captures some of the more important ones.

Writeback from cache to DRAM (or nontemporal write):

• from a single core with no contention or sharing: ** Nontemporal cacheline writes issue at nearly 123M/s; read-for-ownership + normal writes seem to be only a bit slower, around 113M/s in order and dropping off to ~100M +- 5M with random addresses.
• Stay in fill buffers until complete. There are 10 fill buffers.
• Thus, it takes about 81ns to complete a writeback in the best case.