Fast ESP (and FSIS) on a SAN RRS feed

  • General discussion

  • I have found a few documents stating the the SAN-based topology would be the best option for a high performance Fast ESP site. I am talking about implementing a multi-node, redundant site, with partitioned indexing. We do have a state-of-the-art IT that already knows how to handle SANs and clustered file systems for several other needs, so having the proper assets and staff is not an issue. The problem is that there is nothing specific on the "Fast Deployment Guide" or on the "Fast Installation Guide" books.

    The Deployment Guide says: SAN offers improved performance compared with the use of local disk or NAS (Network Attached Storage). Combined with the use of a cluster filesystem, the performance is in general further improved due to the elimination of index copy time in multi-node configurations”.

    The whole smell of the thing makes me think that SAN topologies were a Unix-only option at the time. Nowadays apparently we have a reliable clustered file system on Windows Server 2008 R2 Enterprise. My IT staff knows how to handle it too. So here we go:

    In order to setup Fast ESP on a SAN topology:

    1. What Windows version do we need at every Fast node? Is it Windows Server 2008 R2 Enterprise?
    2. Do we need to mount the clustered file system at every node? Or is it done only at the Indexing nodes? What about the Search nodes?
    3. I was told by the local MS partner that I should mount a different logical unit at every indexing node (therefore *not* sharing the same clustered file system). That sounds incredibly silly - it would be the same as working with local file systems (logically speaking), the worst performance topology (according to the Deployment Guide). Should I go for a common clustered file system mounted at the nodes where it is needed? Or should I give up using SAN at all?

    Thanks in advance.


    Wednesday, September 15, 2010 6:24 PM

All replies

  • Hi,

    While I am certainly no expert in configuring SAN, i do have experience deploying ESP where SAN has been used for index and fixml. In general i would say that your success installing on SAN depends on two things:

    1) Total size of indexable content

    The less content you have, the greater the probability that SAN will work out all right. I would say that once you get more than 1 000 000 docs per node, you will start to see SAN related performance issues that needs addressing. 

    2) Your SAN guy/people _really_ knows what they're doing.

    That last one might sound silly, given that you just stated that your people knows how to handle SAN's, but in my experience is rare that applications with the intense IO needs of ESP is deployed on SAN, even in larger organizations. 

    You need to mount the clustered file system at all indexing(for the 'data_fixml' and 'data_index' folders) and search nodes (to store the 'data_index' folders).

    At least one project i participated in had to give up SAN before deploying to prod due to performance issues and cost. Though more inconvenient to maintain, it's difficult to beat the raw performance of a local RAID-5 array. 

    Not sure if this was helpful, but this is my 2 cents anyway :)

    - okms
    Monday, October 25, 2010 9:45 PM
  • I can add my two cents to Ole Kristiand's.

    There is one SAN-based installation that I know of that works. It was a 16-node win setup, if I recall correctly. They spent a lot of time tuning the SAN and discovered that the best results came when each server more or less was tied to its own disk in the SAN. (My technical knowledge of SAN is limited). It was described as using local disks, only storing those disks in the SAN.

    A problem with another SAN-based solution was that heavy feeding on the test-environment had a big performance impact on the production-env because of I/O to the same disks. They ended up with local disks.
    Friday, October 29, 2010 9:22 AM