Re: [rdiff-backup-users] What are the bottlenecks in --verify? (Or how to speed up verification?)

[Thomas Harold]

On 10/3/2013 9:52 AM, Dominic Raferd wrote:
> There have been discussions about verification speeds and issues here
> before. I think you are right that the CPU core is a bottleneck as
> rdiff-backup only uses one core when running a verification.
>
> Verifications do use temporary space and can use a lot of it even though
> the temporary files never seem to be visible in the filesystem. I would
> advise an explicit --tempdir setting, and a different spindle with lots
> of space (and speed) would be ideal.

Yeah, I think the verifications definitely can benefit from having a 
spindle dedicated to /tmp (or --tempdir setting).  We added a 2nd set of 
small fast spindles, moved /tmp over, and it seems to have dramatically 
sped up the verify speed.  With "atop" we can see rates of 80-120 MB/s 
being written to the /tmp file system now that it is on its own spindle. 
 So it's getting a lot of use.

From what I can tell, we're still limited by the spindle speeds of both 
the place where the rdiff-backup directory is stored, and also the speed 
of the /tmp (or --tempdir) location.

Our Opteron 6348 (12-core 2.8GHz) doesn't seem to have trouble 
calculating the SHA1 hashes that rdiff-backup uses.

I don't have hard numbers on rdiff-backup itself, but our script to 
rsync the rdiff-backup files off to an external USB3 drive now takes 
about 12 hours instead of 18 hours.  It also does --verify of the source 
directory prior to the rsync, then does a --verify of the target 
directory after the rsync.  I have to examine things again to see if I 
can cut that time window down to about 4-5 hours.

The next step for us will be to move the rdiff-backup directories to a 
PV with more and faster spindles.