| > Hi there,
Hi.
> Suppose the following ( default example from HSZ manual ):
> add disk disk.....
> add mirrorset m1 disk100 disk200
> add mirrorset m2 disk300 disk400
> add mirrorset m3 disk500 disk600
> add stripeset s1 m1 m2 m3
> add unit ....
> When data is coming, what exactly happens here ?
> a) data is striped to disk100 disk300 disk500 and mirrored to
> disk200 disk400 disk600 afterwards or
> b) is striped across ALL disks and mirrored blockwise afterwards
> in some more complicated algorithm.
a) is closest to what happens, but not quite. The data is striped across
the three mirrorsets, and then writes are *simultaneously* issued to both
mirrorset members. It is not a case of "write one, wait, write the other".
Rather it is "issue writes to both, wait for both to complete".
> The reason I ask is: if I will use bigger stripes, I need to define
> an order ( in case a ) like this:
> add mirrorset m1 disk100 disk200
> ...
> add mirrorset m4 disk210 disk110
> to get a striping across all ports ( first member of m1 is disk100,
> first member of m4 is disk210 - on second port ).
Given the above, going through the trouble of specifying port order carefully
when you define the mirrorsets will not make any difference
on writes. It will, however, help to level the port load on reads if the
read selection algorithm is LEAST_BUSY. See the discussion for 1192 in
HSJ40_PRODUCT conference.
Bob Bean
Array Controller Engineering
|