[DGD] minimum sector size tweaking

Raymond Jennings shentino at gmail.com
Fri May 17 01:30:41 CEST 2013


I've completed my analysis.

It appears that there is a fixed length structure at the head of the
snapshot occupying 64 bytes.  Combined with the other header information
present which occupies only 12 bytes.

Control and dataspace block header sizes are comfortably under 32 bytes.

Therefore, 128 bytes appears to be the smallest power of two permissible as
a sector size without causing data corruption problems.

It's debatable as to whether or not anything smaller than 512 bytes is
efficient, but to those for whom disk space is at a premium it appears that
having a sector size of 128 bytes (edit config.c to override the limit) is
amply safe.

It's not (yet?) officially allowed in the mainline though, so be wary of
breaking compatibility.




On Mon, Apr 29, 2013 at 3:26 AM, Raymond Jennings <shentino at gmail.com>wrote:

> I presume that it depends on the host platform because it depends on data
> type sizes of the header's members?
>
> So far I've found:
>
> dataspace and control block headers at the beginning of a normal sector
> signature and dumpfile headers at the head and tail of the "boot" sector
>
> Did I miss anything?
>
> In the meantime I've noticed respectable savings in disk usage.
>
>
> On Mon, Apr 29, 2013 at 3:10 AM, Felix A. Croes <felix at dworkin.nl> wrote:
>
>> Raymond Jennings <shentino at gmail.com> wrote:
>>
>> > I've been tinkering with the minimum sector size setting in config.c and
>> > bumped it down to 128 for now.
>> >
>> > I'm aware of some header information that is confined to the first
>> sector,
>> > which is why I didn't shrink it much beyond that.
>> >
>> > How small is safe?  I'm thinking 512 is a lower bound for efficiency
>> > reasons, but is it actually dangerous to go below that?
>>
>> It depends on the host platform.  You can use binary search to find out.
>>
>> Regards,
>> Felix Croes
>> ____________________________________________
>> https://mail.dworkin.nl/mailman/listinfo/dgd
>>
>
>



More information about the DGD mailing list