DragonFly BSD
DragonFly kernel List (threaded) for 2011-05
[Date Prev][Date Next]  [Thread Prev][Thread Next]  [Date Index][Thread Index]

Re: HAMMER2 design in progress - 1-2 year time frame


From: Alex Burke <alexjeffburke@xxxxxxxxx>
Date: Mon, 23 May 2011 10:24:46 +0100

Hi,

Didn't get the chance to say thank you for your reply - found it very
interesting and it definitely explained a few things.

It definitely think DragonFly is doing unique things with some very
novel approaches; best of luck with the implementation once it begins.

Thanks, Alex J Burke.

On 13 May 2011 16:10, Matthew Dillon <dillon@apollo.backplane.com> wrote:
>
> :Hi,
> :
> :I'm a DFBSD lurker and I certainly wouldn't profess to understand
> :everything in that document, but a comp sci degree means I just about
> :followed :)
> :
> :I noticed you had decided not to store the versioned inodes in the
> :B-Tree, which I believe is what HAMMER 1 did. I was curious why that
> :was - I saw mention of not being able to have parent pointers in on
> :media structures, but want sure if it was related and what the impact
> :was.
> :
> :Otherwise, sounds pretty incredible; that any sub trees can be PFS,
> :writable snapshots and multiple roots are all ingenious.
> :
> :Cheers, Alex J Burke.
>
>    The B-Tree is a great indexing mechanism but it suffers from
>    two big problems in HAMMER1.
>
>    First it is very difficult to keep things organized for locality of
>    reference.  For example when you split a B-Tree node, the split node
>    can wind up seriously far away from the original with regards to
>    disk seek times.  The seek inefficiencies are largely removed if one
>    uses swapcache.
>
>    Second and more importantly, manipulation of the B-Tree requires
>    an UNDO FIFO (which HAMMER1 implements).  The UNDO FIFO is one of
>    HAMMER1's most complex bits of code and how other implications as
>    well, such as having to hold dirty buffers locked to prevent the
>    kernel from writing them out too early (the UNDO data has to be
>    flushed first).
>
>    HAMMER2 primarily addresses the second point.  By chasing blocks
>    all the way to the root HAMMER2 doesn't need an UNDO FIFO at all,
>    and (theoretically) doesn't need special handling of the buffer
>    cache by the OS because write ordering no longer matters.  The
>    disk flush sequence prior to updating the volume root is still
>    necessary but it's still far less complex than HAMMER1's UNDO
>    sequencing.
>
>    The trade-off is that HAMMER2 currently does not index its inode space
>    by inode number (there being no B-Tree or other data structure to
>    maintain such an index in), which is its primary reason for not being
>    able to support hardlinks.  On the plus side the B-Tree in HAMMER1
>    could support only simple snapshots while the block referential model
>    HAMMER2 will use can also support writable snapshots (i.e. snapshot
>    forks).
>
>                                                -Matt
>




[Date Prev][Date Next]  [Thread Prev][Thread Next]  [Date Index][Thread Index]