File Systems

From Siwiki

Contents 1 ZFS 2 NFS 2.1 Client Performance Rules of Thumb 2.2 Optimize networking 2.3 NFS Client: Open/Close optimization 3 UFS 4 QFS 5 Segmap 6 DTrace

[edit] ZFS

• ZFS Best Practices Guide, Under Construction
• ZFS Configuration Guide, Under Construction
• ZFS Performance Guide, Under Construction
• ZFS for Databases Guide, Under Construction
• The Dynamics of ZFS
• Tuning ZFS Record Size
• ZFS Ram Usage
• When (and when not to) use RAID-Z
• ZFS and Direct I/O
• Neel's ZFS for Databases Blog

[edit] NFS

Other References

• Using NFS with Direct I/O for Oracle Databases
• NFS Server Tuning

[edit] Client Performance Rules of Thumb

• Sequential performance is making great strides
• Performance Rules of thumb
• NFS3 on 1GHz SPARC, Solaris 8: 30MB/s
• NFS3 on 1GHz SPARC, Solaris 9: 55MB/s
• NFS4 on 1GHz SPARC, Solaris 10: 90MB/s
• NFS4 on 2Ghz Opteron, Solaris 10: 105MB/s (wirespeed!)
• NFS4 on 2Ghz Opteron, Solaris 10, 10Gbit ethernet: 255MB/s

[edit] Optimize networking

• Use Gigabit networks (10GBE is avail. Now too...)
• Use Jumbo-frames where possible
• Dedicate a network to storage (Just like we do with a SAN)
• Maximize NFS block size
  • Solaris 8 has a max of 32k
  • Solaris 9 allows up to 1MB! (Solaris 9 NFS Server required)
• Set nfs3_bsize and nfs3_max_transfer_size system parameter on client and server
  • Futher tuning down of blocksize can be done via mount options: rsize, wsize

[edit] NFS Client: Open/Close optimization

• NFS uses open-to-close semantics to attempt to optimize consistency between clients
  • Drawback is wait on close for flush...
• NFS client has a mount option to optimize for this condition
  • Open/close intensive applications can be improved
  • Useful only if no sharing is occuring

# mount -o nocto ...

[edit] UFS

• Using Direct I/O to provide raw-disk like performance
• The filestat utility

[edit] QFS

• Using Direct I/O to provide raw-disk like performance

[edit] Segmap

• Segmap Buffer Cache Tuning

[edit] DTrace

• Using DTrace to analyze page cache hit rates