> Cpu usage is 4x lower than librbd, so it's possible to reach a lot more iops (I have > benched 100000 iops with mutiple iothreads). But this is a bug inside librbd, so they should fix it there? And this does not work with our 2.6.32 kernel...