Francois Tigeot benchmarked the recent Postgres 9.3 release. Postgres apparently switched to using mmap instead of SYSV shared memory, and Francois has done this to show the performance differences. (view the PDF in his post.) Of course, work has continued since this was posted, so there should be new numbers soon, and new changes I’ll document in a future post.
I haven’t found a reference to the exact decision Postgres made on how to handle memory; please post a link in comments if you know a good source.
http://rhaas.blogspot.nl/2012/06/absurd-shared-memory-limits.html