Just to consider all possibilities...
Current architecture needs to create all directory structure on all
bricks, and has the big problem that each directory in each brick will
store the files in different order and with different d_off values.
This is a serious scalability issue and have many inconveniences when
trying to heal or detect inconsistencies between bricks (basically we
would need to read full directory contents of each brick to compare them).
An alternative would be to convert directories into regular files from
the brick point of view.
The benefits of this would be:
* d_off would be controlled by gluster, so all bricks would have the
same d_off and order. No need to use any d_off mapping or transformation.
* Directories could take advantage of replication and disperse self-heal
procedures. They could be treated as files and be healed more easily. A
corrupted brick would not produce invalid directory contents, and file
duplication in directory listing would be avoided.
* Many of the complexities in DHT, AFR and EC to manage directories
would be removed.
The main issue could be the need of an upper level xlator that would
transform directory requests into file modifications and would be
responsible of managing all d_off assignment and directory manipulation
(renames, links, unlinks, ...).
Xavi
On 12/16/2014 03:06 AM, Anand Avati wrote:
Replies inline
On Mon Dec 15 2014 at 12:46:41 PM Shyam <srangana@xxxxxxxxxx
<mailto:srangana@xxxxxxxxxx>> wrote:
With the changes present in [1] and [2],
A short explanation of the change would be, we encode the subvol ID in
the d_off, losing 'n + 1' bits in case the high order n+1 bits of the
underlying xlator returned d_off is not free. (Best to read the commit
message for [1] :) )
Although not related to the latest patch, here is something to consider
for the future:
We now have DHT, AFR, EC(?), DHT over DHT (Tier) which need subvol
encoding in the returned readdir offset. Due to this, the loss in bits
_may_ cause unwanted offset behavior, when used in the current scheme.
As we would end up eating more bits than what we do at present.
Or IOW, we could be invalidating the assumption "both EXT4/XFS are
tolerant in terms of the accuracy of the value presented
back in seekdir().
XFS has not been a problem, since it always returns 32bit d_off. With
Ext4, it has been noted that it is tolerant to sacrificing the lower
bits in accuracy.
i.e, a seekdir(val) actually seeks to the entry which
has the "closest" true offset."
Should we reconsider an in memory _cookie_ like approach that can help
in this case?
It would invalidate (some or all based on the implementation) the
following constraints that the current design resolves, (from, [1])
- Nothing to "remember in memory" or evict "old entries".
- Works fine across NFS server reboots and also NFS head failover.
- Tolerant to seekdir() to arbitrary locations.
But, would provide a more reliable readdir offset for use (when valid
and not evicted, say).
How would NFS adapt to this? Does Ganesha need a better scheme when
doing multi-head NFS fail over?
Ganesha just offloads the responsibility to the FSAL layer to give
stable dir cookies (as it rightly should)
Thoughts?
I think we need to analyze the actual assumption/problem here.
Remembering things in memory comes with the limitations you note above,
and may after all, still not be necessary. Let's look at the two
approaches taken:
- Small backend offsets: like XFS, the offsets fit in 32bits, and we are
left with another 32bits of freedom to encode what we want. There is no
problem here until our nested encoding requirements cross 32bits of
space. So let's ignore this for now.
- Large backend offsets: Ext4 being the primary target. Here we observe
that the backend filesystem is tolerant to sacrificing the accuracy of
lower bits. So we overwrite the lower bits with our subvolume encoding
information, and the number of bits used to encode is implicit in the
subvolume cardinality of that translator. While this works fine with a
single transformation, it is clearly a problem when the transformation
is nested with the same algorithm. The reason is quite simple: while the
lower bits were disposable when the cookie was taken fresh from Ext4,
once transformed the same lower bits are now "holy" and cannot be
overwritten carelessly, at least without dire consequences. The higher
level xlators need to take up the "next higher bits", past the previous
transformation boundary, to encode the next subvolume information. Once
the d_off transformation algorithms are fixed to give such due "respect"
to the lower layer's transformation and use a different real estate, we
might actually notice that the problem may not need such a deep redesign
after all.
Hope that helps
Thanks
Shyam
[1] http://review.gluster.org/#/c/__4711/
<http://review.gluster.org/#/c/4711/>
[2] http://review.gluster.org/#/c/__8201/
<http://review.gluster.org/#/c/8201/>
_________________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx <mailto:Gluster-devel@xxxxxxxxxxx>
http://supercolony.gluster.__org/mailman/listinfo/gluster-__devel
<http://supercolony.gluster.org/mailman/listinfo/gluster-devel>
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
http://supercolony.gluster.org/mailman/listinfo/gluster-devel
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
http://supercolony.gluster.org/mailman/listinfo/gluster-devel
