4 Replies Latest reply: Feb 1, 2010 4:54 AM by Billy~Verreynne RSS

    Oracle on zfs

    tolik2525
      Hello!
      We have Solaris 10 box with latest patches connected to Storedge 6130 by FC.
      Storedge disks configured as RAID5
      Our Oracle 10g (10.2.0.4) DB performance is terrible.

      Write speed reported by "zpool iostat" about 5 MB/s and "iostat" showing 100% load for data pool.
      Redo logs and data files placed on different pools. Recordsize for log - 128k, for data files 8k.
      I have no idea what else i can tune. There are guides for tuning zfs for oracle (best practices guide and evil tuning guide) but nothing for tuning oracle for zfs.

      I'm also worried about fact that simple
      dd if=/dev/zero of=/zpool1 bs=8192k count=10240
      gives 40MB/s in zpool iostat

      Maybe Oracle settings are wrong?
      Current settings are:
      filesystemio_options=asynch
      db_writer_processes=2
      db_file_multiblock_read_count=8
      Does ZFS support DirectIO and/or asynchronous writes like UFS?

      Please help.
      Thanks in advance!
      (and sorry for my English)
        • 1. Re: Oracle on zfs
          Billy~Verreynne
          tolik2525 wrote:

          I'm also worried about fact that simple
          dd if=/dev/zero of=/zpool1 bs=8192k count=10240
          gives 40MB/s in zpool iostat
          This means that I/O itself is simply slow.
          Maybe Oracle settings are wrong?
          How can Oracle be at fault when your diagnosis of the underlying file system shows it to perform poorly? How is Oracle support to rectify the storage system's poor performance?


          Note that RAID5 requires a parity calculation for every single write() to that storage system. If that calculation is not done asynchronously via something like ASICs (Application Specific Integrated Circuits), it means that the I/O latency includes the parity calculation.

          Obviously this will severely impact performance. (search for BAARF via your favourite search engine for more details).
          • 2. Re: Oracle on zfs
            tolik2525
            I agree that RAID5 is bad and my storage is slow.

            But Oracle definitely involved.
            From Oracle side for UFS i have clear instructions: "use directio, and set filesystemio_options=SETALL". For ZFS i have nothing.
            Trying to change parameters and watch what has been affected is not an option. Settings differ from one zfs version to another and in combination with oracle parameters it's almost unreal to find right combinations without some sort of manual.
            • 3. Re: Oracle on zfs
              ora_tech
              Hi,

              I Suggest you to use oracle ASM technology for your database.

              Refer link below:

              http://www.oracle.com/technology/products/database/asm/index.html

              Regards,
              X A H E E R
              • 4. Re: Oracle on zfs
                Billy~Verreynne
                tolik2525 wrote:
                I agree that RAID5 is bad and my storage is slow.

                But Oracle definitely involved.
                From Oracle side for UFS i have clear instructions: "use directio, and set filesystemio_options=SETALL". For ZFS i have nothing.
                Trying to change parameters and watch what has been affected is not an option. Settings differ from one zfs version to another and in combination with oracle parameters it's almost unreal to find right combinations without some sort of manual.
                So from that it would seem that a) that no specific parameters need to be set, or b) that zfs is not exactly a supported and recommended file system.

                In which case, if Metalink/MOS notes lack sufficient info, file a SR to get clarification from Oracle Support.

                But no amount of direct and async I/O will alleviate poor I/O performance of the storage system itself.