It appears that the impact of buffered vs. memory-mapped index file reads has changed dramatically since last I tested. Here's some results on various platforms we pulled together yesterday w/2.2-HEAD.
TL;DR: On linux we see a 40% hit in performance from 108k ops/sec on reads to 64.8k ops/sec. While surprising in itself, the really unexpected result (to me) is on Windows - with standard access we're getting 16.8k ops/second on our bare-metal perf boxes vs. 184.7k ops/sec with memory-mapped index files, an over 10-fold increase in throughput. While testing w/standard access, CPU's on the stress machine and C* node are both sitting < 4%, network doesn't appear bottlenecked, resource monitor doesn't show anything interesting, and performance counters in the kernel show very little. Changes in thread count simply serve to increase median latency w/out impacting any other visible metric that we're measuring, so I'm at a loss as to why the disparity is so huge on the platform.
The combination of my changes to get the 2.1 branch to behave on Windows along with benedict and Stefania's changes in lifecycle and cleanup patterns on 2.2 should hopefully have us in a state where transitioning back to using memory-mapped I/O on Windows will only cause trouble on snapshot deletion. Fairly simple runs of stress w/compaction aren't popping up any obvious errors on file access or renaming - I'm going to do some much heavier testing (ccm multi-node clusters, long stress w/repair and compaction, etc) and see if there's any outstanding issues that need to be stamped out to call mmap'ed index files on Windows safe. The one thing we'll never be able to support is deletion of snapshots while a node is running and sstables are mapped, but for a > 10x throughput increase I think users would be willing to make that sacrifice.
The combination of the powercfg profile change, the kernel timer resolution, and memory-mapped index files are giving some pretty interesting performance numbers on EC2.