On Tue, Jul 19, 2011 at 12:45:55AM -0400, Venkatesh Srinivas wrote: > Could you give blogbench a try with vm.randomize_mmap = 1? I don't > think the I/O performance will change, but I think you will see less > contention on vm_token. (mmap times will be shorter...)
I tried it briefly, performance results were the same. I saw almost no threads in vm_token state with these last tests; it only happened when there was much more memory which could be used for caches (8GB or more). -- Francois Tigeot
