On mer, 2005-04-27 at 15:42 -0400, goggin, edward wrote: > ... especially for SCSI paths where the multipathing is designed to be > indifferent to both transport and network route between SCSI initiator > and SCSI target? Don't they both result in one path per path group > for a given SCSI logical unit as long as the node name is defined to > be the "host number/channel number/target id" triple setup by > sysfs_devinfo in libmultipath/discovery.c? > > BTW, this is an example of the SCSI affinity I was mentioning in my > previous submission. > > -- Tunables don't account as blocking affinity : if you want to multipath over ieee1394, you just won't use those tunables like group-by-nodename. The fact is, the world seems full of exotic hardware. This pgpolicy was asked by IBM for their ESS class controllers were each node has a unique nodename whatever the number of paths to the node. Having a penalty switching a LU from one node to its peer, they were interested in group all paths to a node together. Well, at least it's what I understood from the request :/ One would have thought that group_by_serial was here to cover that scenario, but that hardware just doesn't work that way. Regards, -- christophe varoqui <christophe.varoqui@xxxxxxx>