Previously the majority of the VM was under a single token, the vm_token; now vm_objects (mappable entities) are each under a private token, concurrent page faults in the same object can proceed, and VM SMP scalability overall improved. Results for Postgres and MySQL, among others, are excellent.
+Francois Tigeot [worked on benchmarks](http://leaf.dragonflybsd.org/mailarchive/kernel/2011-11/msg00004.html) comparing DragonFly 2.10 and DragonFly 2.13, among other operating systems. The [resulting graphs](http://dl.wolfpond.org/Pg-benchmarks.pdf) show how well DragonFly responds to multiple CPUs. (2.13 is the version of DragonFly that became 3.0.)
See:
-http://leaf.dragonflybsd.org/mailarchive/kernel/2011-11/msg00004.html
-http://dl.wolfpond.org/Pg-benchmarks.pdf
-### HAMMER performance greatly increased
-A new time domain multiplexing method has been added to balance storage operation types over long time periods. As a result, HAMMER performance increased by XX% under some typical scenarios (see benchmarks here, here, here) and HAMMER fairness and throughput under heavy storage loads is now better than all tested alternatives.
+### HAMMER performance greatly increased
-http://leaf.dragonflybsd.org/mailarchive/commits/2011-07/msg00086.html
+A new [time domain multiplexing method](http://leaf.dragonflybsd.org/mailarchive/commits/2011-07/msg00086.html) has been added to balance storage operation types over long time periods. HAMMER fairness and throughput under heavy storage loads has significantly improved.
-HAMMER's behavior in very low memory configurations has also improved; it will no longer readily starve a system of kernel allocation space.
+HAMMER's behavior in very low memory configurations has also improved; it will no longer readily starve a system of kernel allocation space. HAMMER should work on systems with as little as 256M of RAM.
### SMP kernel installed by default