Topic
  • 1 reply
  • Latest Post - ‏2013-12-04T23:49:18Z by yuri
JGS723
JGS723
1 Post

Pinned topic Optimized File Transfer to SONAS

‏2013-12-04T20:46:55Z |

Client is working on finding the absolute most efficient process possible for transferring 100s of terabytes of data into the SONAS from across campus.  The 10G link is in place, and max thru put is showing high numbers for short transfers, but sustained rates are what they really want to optimize.   Client is looking for suggestions.  The questions they are asking are,  "Can we connect directly to an interface node?" , "Can you connect to a storage node? and is that faster?" and, "Is there's a udp option that gives a performance boost, that's the kind of info we'd like.
 

Any suggestions will be much appreciated.

JGS - jslivinski@microstrat.com
 

  • yuri
    yuri
    280 Posts

    Re: Optimized File Transfer to SONAS

    ‏2013-12-04T23:49:18Z  

    Going around standard SoNAS processes is unlikely to bring happiness, but is guaranteed to make support more difficult.  

    The key to getting good overall performance is parallelism.  Subdividing the dataset and using as many, say, rsync threads as practical will raise the overall throughput rate.  I'm not sure what "connect directly" means; presumably the intent here is to avoid using NFS, but in favor of what protocol?  Instead of trying to bypass NFS, I would recommend applying all of the recommended NFS tuning (in particular, using large wsize).  In any event, don't try to do IO on storage nodes -- those are not configured to handle IO directly at high rates (caches are very small on those nodes), and this doesn't eliminate the need to do network IO, since only a subset of disks is locally visible on a given storage node.

    yuri