2011-11-04

Solaris: Poor IO performance on Veritas for 11.2 compared to 10.2

This problem is introduced in RDBMS 11.2.0.1 on Solaris.
 
The IO code in 11.2 disables the VXFS cache by issuing
 ioctl(fd, VX_SETCACHE, VX_DIRECT).
The fix re-enables VXFS cache based on settings of 
filesystemio_options = asynch | none but only if cluster_database 
is not set to TRUE (as VXFS caching should not be used with RAC).
 
Rediscovery Notes:
 You may be seeing this problem if:
  There is a considerable IO slow down compared to 10.2.x 
  (eg: by a factor of 5x) for IO events like 
  "db file sequential read" and "db file parallel read"
 AND
  Files are located on VXFS with VXFS cache enabled 
  (ie. VRTSodm is not used and mount options mincache=direct,
   convosync=direct are not set)
 
Workaround
 Increase db_cache_size parameter.
 
 If RAC is experiencing this problem after 10g upgrade then there is no other
 solution than increasing the db cache.
 
Note:
 The behavior of this fix is on Solaris diverges from HP-UX
 where VXFS_DIRECT is done only based on filesystemio_options 
 settings of directio | setall but independent of RAC mode being 
 on or not.
 

Niciun comentariu:

Trimiteți un comentariu