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...) Thanks, -- vs