6.3.1.4 What to Look For When Monitoring Exadata Smart Flash Cache

Increased Read Latencies

Possible issues related to Exadata Smart Flash Cache tend to be visible in the database as increased read latencies, specifically in the cell single block physical read wait event. The increased latency is usually caused when reads are issued against the hard disks instead of being satisfied by Exadata Smart Flash Cache.

Read requests may not be satisfied from Exadata Smart Flash Cache when:

• The required data is not in Exadata Smart Flash Cache.

  In this case, requests are recorded as flash cache misses, which are visible as increased misses in the Flash Cache User Reads section of the AWR report. This is also typically accompanied by increased population writes, which are visible in the Flash Cache Internal Writes section of the AWR report.

  On the database, you may see a reduction in the number of cell flash cache read hits compared with the physical read IO requests or physical read total IO requests.

• The data is not eligible for Exadata Smart Flash Cache.

  To maximize cache efficiency, when an I/O request is sent from the database to the storage servers, it includes a hint that indicates whether or not the data should be cached in Exadata Smart Flash Cache.

  If data is not eligible for caching, the corresponding I/O is visible in the Flash Cache User Reads - Skips section of the AWR report, along with the reason why the read was not eligible. Possible reasons include:

  • The grid disk caching policy is set to none.
  • The database segment is configured with the CELL_FLASH_CACHE NONE storage option.
  • The I/O type and situation precludes caching. For example, the I/O is related to an RMAN backup operation and the hard disks are not busy. If this is a common occurrence, it will be evident in Top IO Reasons section of the AWR report

In some cases, the cell single block physical read latency may increase, with no apparent difference in Exadata Smart Flash Cache behavior. This can be caused by increased I/O load, especially on hard disks.

Occasional long latencies in cell single block physical read, usually visible as a long tail in the cell single block physical read histogram may simply indicate that the data is read from hard disk, and does not necessarily indicate a problem with Exadata Smart Flash Cache.

Skipped Writes

When using Exadata Smart Flash Cache in Write-Back mode, most database write requests are absorbed by the cache. However, some write requests may skip the cache. The most common reason for skipping the cache is when the request writes a large amount of data that will be read once, such as temp sorts, or it is not expected to be read at all in the foreseeable future, such as backups and archiving.

Another common reason for skipping large writes is Disk Not Busy, which typically means that there is no benefit to using Exadata Smart Flash Cache because the hard disks have sufficient capacity to handle the write requests.

If skipping Exadata Smart Flash Cache for large writes is causing a performance issue, then it is typically visible in the database with corresponding long latencies for the direct path write or direct path write temp wait events.

The reasons for rejecting large writes are visible in the Flash Cache User Writes - Large Write Rejections section of the AWR report.

Database Working Set Size

The database working set refers to the subset of most commonly accessed information in the database. In most cases, the database working set is fairly stable. However, if for any reason, the working set does not fit in Exadata Smart Flash Cache, you may see the following symptoms:

• Increased cache misses, indicating that data is not in Exadata Smart Flash Cache. This is visible in the Flash Cache User Reads section of the AWR report, or the FC_IO_RQ_R_MISS_SEC cell metric.
• Increased population activity to populate data not in Exadata Smart Flash Cache. This is visible in the Flash Cache Internal Writes section of the AWR report, or the FC_IO_[BY|RQ]_W_POPULATE_SEC cell metrics.
• Increased disk writer activity, indicating that dirty cachelines have to be written to disk, so that the cacheline can be reused to cache other data. Disk writer activity is visible in the Flash Cache Internal Reads section of the AWR report, or the FC_IO_[RQ|BY]_[R|W]_DISK_WRITER_SEC cell metrics.
• Increased first writes, indicating new data is being written to Exadata Smart Flash Cache. A large number of first writes with few overwrites means new data is being written to Exadata Smart Flash Cache. This is visible in the Flash Cache User Writes section of the AWR report, or in the FC_IO_[RQ|BY]_W_FIRST_SEC cell metrics.

In this case:

• Review the database access patterns for tuning opportunities that reduce the amount of data being accessed.
• Consider increasing the number of available storage servers to deliver more space for Exadata Smart Flash Cache.
• Review your I/O Resource Management (IORM) quotas for Exadata Smart Flash Cache and allocate space where it is most required.
• Consider using Extreme Flash storage servers to eliminate the disk I/Os.

Flash Cache Size Advisor

Oracle Exadata System Software release 26.2.0 introduces Flash Cache Size Advisor. The advisor works in conjunction with Oracle AI Database version 23.26.4 (October 2026 Release Update) or later.

The advisor runs on the Exadata storage servers, where it analyzes workload I/O activity over time and automatically generates recommendations for refining Exadata Smart Flash Cache allocations. Advisory information from multiple storage servers is collected in the Oracle AI Database Automatic Workload Repository (AWR), and recommendations are aggregated and delivered through the Automatic Database Diagnostic Monitor (ADDM) framework. You can view advisor recommendations in ADDM reports or in the ADDM section of AWR reports.

Look for recommendations when indicators such as cache misses, population writes, and disk-writer activity suggest that the database working set may not fit in flash cache. Each recommendation includes an estimated benefit and a summary of its rationale. When IORM manages flash cache allocations for individual databases, the advisor can recommend increasing allocations for specific databases. When flash cache resources are not limited for individual databases, it can recommend adding storage servers to increase overall flash cache capacity. Reassess recommendations over time to evaluate the effects of changes and to account for workload changes.

Miscellaneous Issues

There are some cases when the increased cell single block physical read latency may not be due to cell performance, but may be caused by something else along the IO path, such as network contention or contention for database server CPU resources.

The histograms for cell single block physical read and small reads are available in the Single Block Reads and Small Read Histogram - Detail sections of the AWR report under Exadata Statistics > Performance Summary. The cell single block physical read histogram shows latencies measured by Oracle Database, while the small read histograms show latencies measured in the storage server.

A histogram with a significant number of occasional long latencies is said to have a long tail. When the histograms for cell single block physical read and small reads have long tails, then this is an indication of slow read times in the storage server, which would warrant further investigation of the other I/O performance statistics. See Monitoring Cell Disk I/O.

If the cell single block physical read histogram has a long tail that is not present in the small read histograms, then the cause is generally not in the storage server, but rather something else in the I/O path, such as bottlenecks in the network or contention for compute node CPU.