If you encounter a situation where you see very poor performance on a Sun Fire 15K domain after migrating an application from a Sun Enterprise 10000 domain, the following case information might prove useful for troubleshooting. You might see similar poor performance on applications that migrate from an UltraSparc[R] II to a UltraSparc[R] III architecture.

Customer's Environment: Sun Fire 15K domain with 17 system boards running the customer's custom application. The application interfaces with an Oracle database. As the number of CPUs (SBs) increases beyond 8 CPUs in a given domain, the performance worsens.

Upon reviewing some performance data, the following was determined:

• `vmstat` showed nothing unexpected: lots of free memory and swap.
• `iostat` showed very little I/O.
• `mpstat` showed high mutex contention (> 20000 smtx/s). This was confirmed with `lockstat`. The biggest culprit seemed to be condition variable usage.
• There was a correspondingly high number of context switches and system calls.

The fix implemented in this case can be found here:

http://soldc.sun.com/articles/oci_cache.html

This greatly improved performance. However, error messages are generated when certain commands are run that touch oraus.msb, but functionality does not appear to be hindered. Contact Oracle to provide the fix for Oracle BUG ID: 2142623, which pertains to this issue.

-- See Radiance Case: 63206282 for details on this specific instance.

-- Contact jd.halverson@sun.com for any problems, changes/additions, discrepancies, etc. with this document.

-- HAS Team Home Page