6.3.8.4 What to Look For When Monitoring I/O Resource Management (IORM)

I/O Latency

Issues with IORM typically result in increased I/O latency. This is usually characterized by higher latency in the cell single block physical read database wait event, and in some cases the cell smart table scan database wait event. If these database wait events are significant, and there is no corresponding latencies in the storage servers that are associated with the flash or disk devices, then this may indicate that IORM is throttling the workload.

To confirm that IORM throttling is occurring:

• In the Top Databases by Requests - Details section of the AWR report, review the Queue Time columns, which show the average amount of time that IORM spent throttling IO requests for the database.
• Review the cell metrics for IORM wait times: DB_IO_WT_SM_RQ, DB_IO_WT_LG_RQ, PDB_IO_WT_SM_RQ, PDB_IO_WT_LG_RQ and CG_IO_WT_SM_RQ.

You can use IORM database statistics and the AWR report to understand the I/O workload as a whole. You can use Exadata metrics to further understand I/O consumption by each category, database, or consumer group. By analyzing statistics and metrics, you can understand which category, database, pluggable database (PDB), or consumer group is not using its resource allocation and which is exceeding its resource allocation.

If the wait times are small or zero, then the plan allocation is sufficient. If the wait times are large, then the plan allocation is insufficient. If the wait times due to high I/O latency result in unacceptable performance, then the IORM plan can be adjusted to give a larger allocation, or additional storage servers may be required to deliver the required I/O resources.

Using IOSTAT

Device utilization and I/O service time statistics monitored using iostat (which is collected by ExaWatcher) are unreliable.

The following is stated in the Linux man page for iostat:

svctm - The average service time (in milliseconds) for I/O requests that were issued to the device. Warning! Do not trust this field any more. This field will be removed in a future sysstat version.

Since the utilization computation depends upon the I/O service time, it is also unreliable.

To monitor the actual I/O utilization for a cell disk, database, pluggable database or consumer group, use the corresponding IORM metrics.