[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

[dm-devel] Needing group_by_target/group_by_node_name


I got dm-multipath working using multibus and roundrobin against a IBM SVC (San Volume Controller) backed SAN. This is instead of using the IBM supplied SDD that only works under 2.4. The SVC is fronting for a number of different IBM/LSILogic RAID controllers. Everything is working fine and failover works (takes 30 seconds but I haven't tried to actually bring tweak the QLA setting so I expect that to go down).

The SVC is actually two linux boxes with 4 FC ports each, all ports in the cluster are active and there is no fail over cost for using paths against different nodes in the cluster, we managed to see over 200 MB/sec sequential write speeds against it, this is inline with expected performance. However the read speed is much slower than on SDD, between 30% and 70% of expected. After some spec reading it seems like even if there is no fail over penalty for switching between SVC cluster node members there is of course a penalty since each SVC has 8GB of cache. So keeping your IO to one controller is a good idea.

However the only way I think I can identify which SVC is the target node_name information (/sys/class/fc_transport/target0\:0\:0/node_name ), so I have to look that up somehow, I am currently unable to figure out how from a block device find the target using /sys.

I assume that once I done this all I need to do is write my own priority finding program that looks these up and groups them into a priority depending on the target node_name /

Thanks for the work on this, we were able to add a second lun, add it to our lvm group and then grow the filesystem on top of it without any downtime!


CTO @ Fotango Ltd

Attachment: PGP.sig
Description: This is a digitally signed message part

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]