For When You Can't Have The Real Thing
[ start | index | login ]
start > NAS


Created by dave. Last edited by dave, 16 years and 337 days ago. Viewed 6,207 times. #4
[diff] [history] [edit] [rdf]

Notes about Network Attached Storage

You have multiple places for problems to occur. Performance areas are:

  • how fast can your disks (or arrays) pass bits
  • how fast can your network devices pass bits
  • how fast can your software translate between the two
General notes:
  • Linux has, in general, a poor nfs server. (This is reflected in our experience; in general, NetApps and Solaris (both SPARC and x86) have been better both in terms of throughput and latency-under-load.)
  • everything else being equal, many small disks will generally be faster than fewer large disks of the same total capacity. (More heads are better.)
  • >>Adaptec's Storage Blog says: a good RAID controller can easily exceed 200MB/s.
  • >>Same place, about Raid-5:
    RAID-5 is can be written in two different ways. If the writes from the host are short, then each write will be converted to two disk reads and two disk writes. Performance will be at BEST 25% of your high water mark. And it will probably be much less than that.

    However if the host writes are long and sequential, then the performance can theoretically approach that of RAID-0. But it’s tough to completely hide the overhead of RAID-5, so getting in the 80% ballpark is pretty good.

    All of these RAID performance numbers are also affected by the controller or OS cache. For example, short sequential RAID-5 writes are typically very slow (because of the four IOs described earlier) but if they can combine in a write-back cache to create long sequential RAID-5 writes, then performance can easily jump 10X.
  • RAID-5 gives excellent read speed (because reads can be stepped across many disks simultaneously) but lousy write speeds (because before a write can be committed the system must confirm the integrity of the RAID stripe, which may require reading the rest of the stripe so a correct parity block can be comitted with the write).
  • RAID-1 gives better write speeds because writes are comitted only twice and don't require extra reads; read speeds are still better than a single disk because they can be stepped across both sides of the mirror at the same time.
  • RAID-0 gives the best read and write performance because all reads and writes are stepped across multiple drives at the same time, however the safety of all the data on the array is at risk of loss due to any single drive failure.
  • NetApp traditionally likes to talk about latency rather than throughput. Our experience has been that NetApps are very good at minimizing latency issues under load.
  • Users sharing NAS will tolerate slow throughput better than they tolerate poor latency -- to a point.
no comments | post comment
This is a collection of techical information, much of it learned the hard way. Consider it a lab book or a /info directory. I doubt much of it will be of use to anyone else.

Useful: | Copyright 2000-2002 Matthias L. Jugel and Stephan J. Schmidt