From: Arnold Hendriks <a.hendriks@b-...> - 2007-02-22 19:19:22
I'm testing a SAN setup with multiple LVM logical volumnes and vblades.
For every lv, I launch a vblade process to offer it to a client.
I notice that when doing heavy I/O to one lv/vbladde, all vblade process
become active. They all seem to have an about equal CPU time. Even
though their combined CPU% is still below 20%, I wonder whether this
might become a scalability problem in the future.
My guess it that the activation of all processes has to do with the AoE
protocol: they probably all have to listen to all incoming traffic just
to pick out the traffic for that specific blade.
Are my assumptions about the problem so far correct?
The only workarounds I see is using partitions instead of separate
vblades (but that would require giving up the 1-to-1 lv-to-vblade-devce
mapping, I think), or assigning a separate MAC address/ethernet device
to each vblade so the kernel will dispatch the packets to the correct
vblade instead of waking them all up (if that is at all possible). As
far as I see, there is no way to have one single vblade host multiple
and separate devices?
Thanks in advance for any information,