Block Device Replication with rdiff

I’ve written a few articles on rdiff-backup, and if you need an increment history to go back in time, rdiff-backup is your tool. 

But what if you just want to replicate a large block device over the Internet? Well, then we turn to the utility that inspired rdiff-backup: rdiff

For our example, we wil assume you are using LVM to create device snapshots—but really, this could be any snapshot or SAN flash implementation. I’ve just written it for Linux’s LVM.

  • /dev/remote-vg0/source will be the device we are replicating from
  • /dev/local-vg0/dest will be the device we are replicating to
  • remotehost is the system that hosts /dev/remote-vg0/source
  • This script is being executed on the destination system.
# Define our source and destination
# (Note: spaces in these paths could break the script)

# Choose a size large enough for the remote write-activity during
# replication

# Must be the same size as $DEST, because rdiff writes in sequential
# order (it thinks the destination is an empty file, so it re-writes
# everything.)
# See Feb 18, 2011 update notes below.  This can be much smaller now if you 
# use the patch below, since writes are avoided unless necessary.

# This is probably safe with the librsync patch discussed below

# Enable compression

# 32k I/O buffers, and 16k blocksize.
RDIFF_OPT='-I 32768 -O 32768 -b 16384 -s'

SOURCE_NAME=`basename "$SOURCE"`

DEST_NAME=`basename "$DEST"`
DEST_SNAP="`dirname $DEST`/$DEST_NAME-snap"

# remove the previous snapshots, if any
ssh $SSHOPTS "$SSHUSER" "lvremove -f '$SOURCE_SNAP'"
lvremove -f "$DEST_SNAP"

# Snapshot the remote host:

# Snapshot the local destination host:

rdiff $RDIFF_OPT -- signature "$DEST_SNAP" - | \
  ssh $SSHOPTS "$SSHUSER" "rdiff $RDIFF_OPT -- delta - '$SOURCE_SNAP' -" | \
  rdiff $RDIFF_OPT -- patch "$DEST_SNAP" - "$DEST"

# Compare the volumes, if you like
md5sum $DEST

# cleanup, remove the snapshots.
ssh $SSHOPTS $SSHUSER "lvremove -f '$SOURCE_SNAP'"
lvremove -f "$DEST_SNAP"

This is a convenient single-pipe process for replication, and it uses the librsync rolling-checksum process, using minimal bandwidth on the network that ssh traverses.

Executing this script yields something like this on a 50GB volume; note that the md5sum’s match perfectly.

  Logical volume "source-snap" created
  Logical volume "dest-snap" created
rdiff: signature statistics: signature[3276800 blocks, 16384 bytes per block]
rdiff: loadsig statistics: signature[3276800 blocks, 16384 bytes per block]
rdiff: delta statistics: literal[27842 cmds, 805715968 bytes, 83492 cmdbytes] copy[1462034 cmds, 52881375232 bytes, 372798 false, 10746386 cmdbytes]
rdiff: patch statistics: literal[27842 cmds, 805715968 bytes, 83492 cmdbytes] copy[1462034 cmds, 52881375232 bytes, 0 false, 10746386 cmdbytes]
7fddc578cdbf5f4e30b7f815e72acebd  /dev/local-vg0/dest
7fddc578cdbf5f4e30b7f815e72acebd  /dev/remote-vg0/source-snap
  Logical volume "source-snap" successfully removed
  Logical volume "dest-snap" successfully removed

Since rdiff does not know the destination is a snapshot of the basis-file, it rewrites the whole thing. Indeed, it could simply seek instead of copy from the basis-file, but the stock rdiff tool does not support this. If I write a patch, it will get posted here—and if you write a patch, please let me know! (see update below!)

Until then, keep double the space free in your volume group that you need to run a snapshot and it should work great!

Update: Fri Feb 18 12:39:51 PST 2011 I just wrote patch for rdiff (within the librsync package) that updates in place, by patching the file-stream-sink code in buf.c. Basically, it reads before writing. If the data read in is the same that it would have written, it skips the write and advances the write pointer; otherwise, it writes as normal. Since this avoids writing to the device that had a snapshot except where necessary, much less snapshot-backing-store is required. This code passes all of the ‘make check’ tests that come with librsync, and I believe it to be stable. On my system, rdiff syncs are about 2x faster due to the much reduced write-overhead of the original implementation.

  • The patch is here
  • and the patched code, ready to compile, is here