This message is from the T13 list server.
Pat,
Power of 2:
I picked power of 2 ratio because it makes the math so easy for modulo
and remainder calculations. Also, it is so easy to tell if you are at the
first LBA of a physical sector, For example if the physical sector size is
4K and if the last three bits of the LBA are 000 you are on the first LBA
of a physical sector. 8K last 4 bits 0000, etc. This makes it very easy
to look for unaligned writes in a trace. I would change it from powers of
2 if there is a practical reason. I haven't heard of one yet.
4K Physical Sector Recommendation:
Since we have to get device driver support for this new feature I thought
it was a good idea to set some starting point. 4K byte sectors seems to be
acceptable as the first jump up from 512 bytes.
Unaligned Write Performance:
I probably should have said "typically" instead of "at least" one rev
penalty for an unaligned write. There will certainly be corner cases
where drives can maintain a stream of unaligned writes by coalescing them
in the buffer and other techniques. I just wanted to point out that
unaligned writes were best avoided.
Best Regards,
DC
Dan Colegrove
Senior Technical Staff Member
Storage Technology Division
IBM Corporation
[EMAIL PROTECTED]
"Pat LaVarre" <[EMAIL PROTECTED]>
Sent by: [EMAIL PROTECTED]
02/15/2002 09:35 AM
To: <[EMAIL PROTECTED]>
cc:
Subject: Re: [t13] Large Physical Sector Proposal Updated
This message is from the T13 list server.
> "Daniel J Colegrove" <[EMAIL PROTECTED]> 02/14/02 05:18PM
> rev. 1 of e01138 Large Physical Sector Support up on the web site.
Fun, thank you.
Delightfully concise: just about two pages of substance. Four pages
appear now in just 17,641 bytes at <http://www.t13.org/docs2002/e01138r1.pdf>.
> allows devices to report
> [a fixed ratio of 1, 2, 4, 8, ...]
> up to 32768 logical sectors per physical sector
Can you explain what motivates limiting the reportable physical/logical
ratio to powers of two?
> existing systems ... restrict practical
> device implementations to 4096 bytes per physical sector.
Yes. I hear we face a 4KiB limit wherever we view disk i/o as a random
track, serial block, access of 4KiB pages of virtual memory.
> 2 Definitions
>
> physical sector: A group of contiguous logical sectors,
> which must be read from or written to the device media
> in a single operation.
>
> logical sector: A uniquely addressable set of 256 words
> (512 bytes).
>
> unaligned write: An unaligned write is a write command
> that does not start at the first logical sector
> of a physical sector or does not end
> at the last logical sector of a physical sector.
Delightfully concise enough to be worth quoting. :-)
> Unaligned Write ... will incur .... penalty
> ... at least a drive revolution.
Yes, though I don't yet have a clear picture how much/little help a
logical sector cache could reasonably offer.
I wonder what kinds of bounded-time guarantees hard drives supply for the
flush of their write caches?
I've heard of low-end designs that guaranteed errors are never deferred
more than one command, but I imagine high-end behaviour is less
determinate?
Pat LaVarre
Subscribe/Unsubscribe instructions can be found at www.t13.org.
Subscribe/Unsubscribe instructions can be found at www.t13.org.
