So in this extreme case avg tps is just 6 transactions better
Great job trying to find the spot where the code worked better. I'm not so sure I trust pgbench results where the TPS was so low though. Which leads us right back to exactly how Jeff measured his original results.
As I said already, I think we need more insight into Jeff's performance report, a way to replicate that test, to look a bit at the latency as reported by the updated LWLock patch that Pierre submitted. Tweaking your test to give more useful results is a nice second opinion on top of that. But we're out of time for now, so this patch is getting returned with feedback. I encourage Jeff to resubmit the same patch or a better one with a little more data on performance measurements to our final 8.5 CommitFest in hopes we can confirm this an improvement worth committing.
Last week I worked on a FUSE based filesystem, which I call BlackholeFS. Its similar to /dev/null, but for directories. Basically it simply returns success for all the writes, but doesn't do any writes on the files under it.
Would moving the pg_xlog/ (and possibly table data too) to such a filesystem exercise this patch better?