We develop a deployment solution that is able to resize up to 16 drives
at a time. While testing, we observed that resize2fs' memory usage
scales with the volume size. Even in case of relatively small file
systems of about 1.5TB, resize2fs would need about 135MB of application
memory. Wit 16 large drives, one would need some GB of RAM just to
resize the drives. Our guess is that this is related to the super block
group bitmaps of ext4 that are held in memory to manipulate/zero their
on-disc counterparts. A quick observation of resize2fs' source revealed
that the latest git tree already has a bitmap interface that allows for
different implementations of the bitmap manipulation algorithms. To
solve the memory usage problem, two solutions seem to be feasible:
- For the huge bitmaps that already exist on the volume, we could create
a disk-backed bitmap implementation that would only cache a small
working set of the entire bitmap. My guess is that we can implement this
efficiently enough to not loose too much performance.
- For the all-zero bitmaps, we could implement a dummy bitmap that
forces all bits to zero and spawns a real bitmap as soon as any bits are
set to one. An alternative would be a tree-based aproach that works
especially well when the bitmap is just sparsely set.
I'd be glad if one of the developers involved in the ext4 development
could tell me if these thoughts make sense and if yes, are there any
plans to incorporate these approaches into the ext library anytime soon
or does it make sense if I would have a deeper look into these issues
and implement them? Thanks in advance for any thoughts about this.
Sincerely
U. Bauer
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
