Re: Improper io_opt setting for md raid5

["Martin K. Petersen" <martin.petersen@oracle.com>]

Damien,

> My bad, yes, that is the definition in sysfs. So io_min is the stride
> size, where:

Depends on the RAID type. For RAID0 and 1 there is no inherent penalty
wrt. writing less than the stride size. But for RAID 5/6 there clearly
is.

> stride size x number of data disks == stripe_size.
>
> Note that chunk_sectors limit is the *stripe* size, not per drive stride.
> Beware of the wording here to avoid confusion (this is all already super
> confusing !).

The choice of "chunk" to describe the LBA boundary queue limit is
unfortunate since MD uses chunk_sectors as the term for what you call
stride.

> Well, at least, that is how I interpret the io_min definition of
> minimum_io_size in Documentation/ABI/stable/sysfs-block. But the wording "For
> RAID arrays it is often the stripe chunk size." is super confusing. Not
> entirely sure if stride or stripe was meant here...

The stripe chunk or stripe unit is what you call stride. Stripe width is
the full stripe across all drives.

> As for read_ahead_kb, it should be bounded by io_opt (upper bound) but
> should be initialized to a smaller value aligned to io_min (if io_opt
> is unreasonably large).

In retrospect I am not really a fan of using io_opt for read_ahead_kb
since, to my knowledge, there is no guarantee that the readahead I/O
will be naturally aligned.

That said, I don't really know of devices where this matters much for
reads. With writes, this would be much more of an issue.

-- 
Martin K. Petersen