[BBLISA] Fwd: Moving 100 GB and 1.3 million files

Rob Taylor rgt at wi.mit.edu
Thu Jul 22 13:57:46 EDT 2010


Hi Ian. This is not really that surprising. Unfortunately, moving large
numbers of small files always seems to have this problem.
As I see it, the problems stem from the time needed for per file
transactional overhead at multiple layers, including the filesystem
,protocol(nfs, cifs), and network(connection setups and teardowns+slow
start). Straight writing of bits is at each level is just plain easier.
A streaming tar using netcat could save some on network connections and
protocol operations, but not on the filesystem. There may be some
filesystem options that you can tweak as well.

I have seen many pieces of software that claim to accelerate small file
transfers, but haven't really found anything that is that great at it.

rgt



On 07/22/2010 01:25 PM, Ian Stokes-Rees wrote:
> 
> 
>  I have a question regarding expectations for file movement between
> disks on adjacent servers.
> 
> Due to a sub-optimal file system layout, I regularly have to move lots
> of files between file systems.  The servers are in the same rack, or at
> least in racks next to each other, and I am fairly certain they are all
> connected to the same GB switch.
> 
> Moving blocks of ~300k files totaling about 5-10 GB takes hours to
> complete.  Yesterday afternoon I started a move of 1.3 million files
> totaling about 94 GB.  20 hours later the transfer seems to be less than
> half done.
> 
> Does this surprise anyone?  Any hints as to what might be wrong or what
> might speed it up?  I'm at a loss to know where to start looking.
> 
> Regards,
> 
> Ian
> 
> 
> 
> More details, for those who are interested:
> 
> The files at the origin are on RAID1 SATA disks (1 TB, ext3, Seagate
> Barracuda 7200 RPM), and I have a ganglia snapshot of the 24 hour status
> (you can see the start of the transfer about 20 hours ago):
> 
> http://dl.dropbox.com/u/1561496/shared/abitibi-origin.pdf
> 
> The destination is an Apple X-RAID array (4TB) connected to an Apple
> XServe.  The corresponding ganglia snapshot is here:
> 
> http://dl.dropbox.com/u/1561496/shared/macintel-destination.pdf
> 
> 
> 
> 
> _______________________________________________
> bblisa mailing list
> bblisa at bblisa.org
> http://www.bblisa.org/mailman/listinfo/bblisa



More information about the bblisa mailing list