On Sat, Oct 08 2016, Dark Penguin wrote:

After researching write-intent bitmaps for a while, my understanding is 
that they are used only to speed up "re-adding" drives by avoiding a 
full resync, and to enable --write-mostly --write-behind. However, it 

This is not correct.  Speeding up re-adding is certainly one benefit.
The other benefit is speeding up resync after a system crash.

does introduce some pretty heavy load on whatever device it's on, 
especially if it's an internal bitmap, because the head would have to 
fly all the way to the superblock twice per each write. If it's an 
external bitmap, then the device it's on would be too busy just serving 
it to do anything else.

It doesn't update the bitmap immediately before and after every write.
Writes are batched, and the bitmap is updated once before each batch of
writes.
There does need to be another update to record that the write has
completed, but that is delayed and usually merged with the update at the
start of the next batch of writes.
So the bitmap is usually updated once per batch of writes.

So if I were to place it on a tmpfs, I could eliminate this problem only 
at the expense of being unable to re-add drives after a reboot, right?.. 
I've read somewhere that bitmaps only work correctly on ext2 or ext3 
filesystems, but that probably means that it's not a good idea to put it 
on a filesystem with delayed allocation like ext4 of zfs, otherwise I 
don't understand why - and so I don't know if there would be any problem 
with it running on tmpfs. Is there?..

You could create a ramdisk, create an ext2 filesystem on that, and put
the bitmap file there.

It probably would make sense to support in-memory bitmaps which never
get written to disk.  It would be fairly easy to do, and would allow
expedited re-add.  I just hasn't been done.
(patches welcome :-)

NeilBrown