I've been trying to convince our DBAs that the backups aren't going to run much faster when they use multiple RMAN streams to backup our databases.

Our config is:

Servers: Sun Servers w/ Solaris 10

Production

CX4-960

Multiple RAID 1/0 LUNs - Each LUN has its own Raid Group.  The LUNs are striped together host side using Oracle ASM.

Backup

CX3-80 - backup disks are locally attached to servers

Multiple RAID5 LUNs - Each LUN has its own Raid Group.  Those LUNs are striped host side using ZFS.

Our DBAs love to use 4 read streams for backups.  I see no improvement over total throughput or total bandwidth.  All I see in fact is increase in production response time, due to increasing queue depth.  For each stream I might see a 5% improvement in bandwidth, but I'll see a 200-300% increase in response time on my production LUNs.  Older servers will actually rob my SP ports of buffer credits because they (the host) can't return R_RDY's fast enough.  I'll actually start to see response time on all LUNs using those same front-end ports increase.  This only happens on older servers (PCI-X bus, not PCIE), newer ones (PCIE bus)  don't seem to have this problem.

I know why this happens.  My problem is:  I'd love to find some Oracle doc, or EMC doc that recommends how many RMAN streams (multiplexing) to use.  I try to convince them that 1 is all they need.  Or some document that says I am wrong - That using lots of streams is the way to go.  I guess I'd just like some good responses with people's experience and how they do things.

I even explain disk thrashing, they seem to think some magic lives inside the CLARiiON to make spinning rust something other than spinning rust.

I might be wrong here, but 4 large, sequential read streams coming from very fast servers doesn't seem very productive.  If we were backing up a bunch of random small files, then I could understand using multiple streams.