Dreadful gmirror performance, though each half works fine

[Andrew Snow]

Pete French wrote:
> I did some benchmarking, and "load" gives me a bit better performance than
> "round-robin" so I've elected to use that. Haven't tried "prefer" as
> syncing all the drives backwards and forwards to get the preferences set
> seems a bit too much like hard work!

I use this patch for sbin/geom/class/mirror/geom_mirror.c

Change:
   md.md_priority = i - 1;
To:
   md.md_priority = i - 1 + 100;

This makes the first disk with a priority of 100 instead of 0, which 
makes it much easier to use "prefer" properly.

> It's frustrating, it is *so* close to being workable with iscsi, and the
> performance is very good, but if it is going to keep locking up on
> me then I just cant use it that way :-(

After failing many times with iSCSI, I use geom_gate with the following 
ggatec options:   -t 30 -q 32768 -R 262144 -S 262144 -o rw

It seems to be very reliable and fast, but you have to use "prefer" to 
get good performance, I only write across the network and not read. 
Load and round-robin lead to slow reads during periods of heavy writes.

Hope that helps,

- Andrew