Actions During Processing

This section discusses some of the details of disk I/O for mirror chunks and how the database server handles media failure for these chunks.

Disk Writes to Mirror Chunks

During database server processing, the database server performs mirroring by executing two writes for each modification: one to the primary chunk and one to the mirror chunk. Virtual processors of the AIO class perform the actual disk I/O. For more information, refer to Asynchronous I/O.

The requesting thread submits the two write requests (one for the primary chunk and one for the mirror chunk) asynchronously. That is, if two AIO virtual processors are idle, they can perform the two disk writes in parallel. In the meantime, the requesting thread can perform any additional processing that does not depend on the result of the mirror I/O.

Disk Reads from Mirror Chunks

The database server uses mirroring to improve read performance because two versions of the data reside on separate disks. A data page is read from either the primary chunk or the mirror chunk, depending on which half of the chunk includes the address of the data page. This feature is called a split read. Split reads improve performance by reducing the disk-seek time. Disk-seek time is reduced because the maximum distance over which the disk head must travel is reduced by half. Figure 102 illustrates a split read.

Figure 102. Split Read Reducing the Maximum Distance Over Which the Disk Head Must Travel

Detection of Media Failures

The database server checks the return code when it first opens a chunk and after any read or write. Whenever the database server detects that a primary (or mirror) chunk device has failed, it sets the chunk-status flag to down (D). For information on chunk-status flags, refer to Mirror Status Flags.

If the database server detects that a primary (or mirror) chunk device has failed, reads and writes continue for the one chunk that remains online. This statement is true even if the administrator intentionally brings down one of the chunks.

Once the administrator recovers the down chunk and returns it to online status, reads are again split between the primary and mirror chunks, and writes are made to both chunks.

Chunk Recovery

The database server uses asynchronous I/O to minimize the time required for recovering a chunk. The read from the chunk that is online can overlap with the write to the down chunk, instead of the two processes occurring serially. That is, the thread that performs the read does not have to wait until the thread that performs the write has finished before it reads more data.