Other information about mmpmon output
When interpreting the results from the mmpmon output there are several points to consider.
Consider
these important points:
- On a node acting as a server of a GPFS™ file system to NFS clients, NFS I/O is accounted for in the statistics. However, the I/O is that which goes between GPFS and NFS. If NFS caches data, in order to achieve better performance, this activity is not recorded.
- I/O requests made at the application level may not be exactly what is reflected to GPFS. This is dependent on the operating system, and other factors. For example, an application read of 100 bytes may result in obtaining, and caching, a 1 MB block of data at a code level above GPFS (such as the libc I/O layer.) . Subsequent reads within this block result in no additional requests to GPFS.
- The counters kept by mmpmon are not atomic and may not be exact in cases of high parallelism or heavy system load. This design minimizes the performance impact associated with gathering statistical data.
- Reads from data cached by GPFS will be reflected in statistics and histogram data. Reads and writes to data cached in software layers above GPFS will be reflected in statistics and histogram data when those layers actually call GPFS for I/O.
- Activity from snapshots affects statistics. I/O activity necessary to maintain a snapshot is counted in the file system statistics.
- Some (generally minor) amount of activity in the root directory of a file system is reflected in the statistics of the file system manager node, and not the node which is running the activity.
- The open count also includes creat() call counts.