15 Process Architecture

The process monitor (PMON) monitors the other background processes and performs process recovery when a server or dispatcher process terminates abnormally.

PMON is responsible for cleaning up the database buffer cache and freeing resources that the client process was using. For example, PMON resets the status of the active transaction table, releases locks that are no longer required, and removes the process ID from the list of active processes.

The listener registration process (LREG) registers information about the database instance and dispatcher processes with the Oracle Net Listener (see "The Oracle Net Listener"). When an instance starts, LREG polls the listener to determine whether it is running. If the listener is running, then LREG passes it relevant parameters. If it is not running, then LREG periodically attempts to contact it.

The system monitor process (SMON) is in charge of a variety of system-level cleanup duties. The duties assigned to SMON include:

• Performing instance recovery, if necessary, at instance startup. In an Oracle RAC database, the SMON process of one database instance can perform instance recovery for a failed instance.

• Recovering terminated transactions that were skipped during instance recovery because of file-read or tablespace offline errors. SMON recovers the transactions when the tablespace or file is brought back online.

• Cleaning up unused temporary segments. For example, Oracle Database allocates extents when creating an index. If the operation fails, then SMON cleans up the temporary space.

• Coalescing contiguous free extents within dictionary-managed tablespaces.

SMON checks regularly to see whether it is needed. Other processes can call SMON if they detect a need for it.

The database writer process (DBW) writes the contents of database buffers to data files. DBW processes write modified buffers in the database buffer cache to disk.

Although one database writer process (DBW0) is adequate for most systems, you can configure additional processes—DBW1 through DBW9, DBWa through DBWz, and BW36 through BW99—to improve write performance if your system modifies data heavily. These additional DBW processes are not useful on uniprocessor systems.

The DBW process writes dirty buffers to disk under the following conditions:

• When a server process cannot find a clean reusable buffer after scanning a threshold number of buffers, it signals DBW to write. DBW writes dirty buffers to disk asynchronously if possible while performing other processing.

• DBW periodically writes buffers to advance the checkpoint, which is the position in the redo thread from which instance recovery begins. The log position of the checkpoint is determined by the oldest dirty buffer in the buffer cache.

In many cases the blocks that DBW writes are scattered throughout the disk. Thus, the writes tend to be slower than the sequential writes performed by LGWR. DBW performs multiblock writes when possible to improve efficiency. The number of blocks written in a multiblock write varies by operating system.

The log writer process (LGWR) manages the online redo log buffer. LGWR writes one portion of the buffer to the online redo log. By separating the tasks of modifying database buffers, performing scattered writes of dirty buffers to disk, and performing fast sequential writes of redo to disk, the database improves performance.

In the following circumstances, LGWR writes all redo entries that have been copied into the buffer since the last time it wrote:

• A user commits a transaction (see "Commits of Transactions").

• An online redo log switch occurs.

• Three seconds have passed since LGWR last wrote.

• The redo log buffer is one-third full or contains 1 MB of buffered data.

• DBW must write modified buffers to disk.

  Before DBW can write a dirty buffer, the database must write to disk the redo records associated with changes to the buffer (the write-ahead protocol). If DBW discovers that some redo records have not been written, it signals LGWR to write the records to disk, and waits for LGWR to complete before writing the data buffers to disk.

The checkpoint process (CKPT) updates the control file and data file headers with checkpoint information and signals DBW to write blocks to disk. Checkpoint information includes the checkpoint position, SCN, location in online redo log to begin recovery, and so on. As shown in Figure 15-4, CKPT does not write data blocks to data files or redo blocks to online redo log files.

Figure 15-4 Checkpoint Process

Description of "Figure 15-4 Checkpoint Process"

The manageability monitor process (MMON) performs many tasks related to the Automatic Workload Repository (AWR). For example, MMON writes when a metric violates its threshold value, taking snapshots, and capturing statistics value for recently modified SQL objects.

The manageability monitor lite process (MMNL) writes statistics from the Active Session History (ASH) buffer in the SGA to disk. MMNL writes to disk when the ASH buffer is full.

In a distributed database, the recoverer process (RECO) automatically resolves failures in distributed transactions.

The RECO process of a node automatically connects to other databases involved in an in-doubt distributed transaction. When RECO reestablishes a connection between the databases, it automatically resolves all in-doubt transactions, removing from each database's pending transaction table any rows that correspond to the resolved transactions.

An archiver process (ARCn) copies online redo log files to offline storage after a redo log switch occurs. These processes can also collect transaction redo data and transmit it to standby database destinations. ARCn processes exist only when the database is in ARCHIVELOG mode and automatic archiving is enabled.

A job queue process runs user jobs, often in batch mode. A job is a user-defined task scheduled to run one or more times. For example, you can use a job queue to schedule a long-running update in the background. Given a start date and a time interval, the job queue processes attempt to run the job at the next occurrence of the interval.

Oracle Database manages job queue processes dynamically, thereby enabling job queue clients to use more job queue processes when required. The database releases resources used by the new processes when they are idle.

Dynamic job queue processes can run many jobs concurrently at a given interval. The sequence of events is as follows:

1. The job coordinator process (CJQ0) is automatically started and stopped as needed by Oracle Scheduler (see "Oracle Scheduler"). The coordinator process periodically selects jobs that need to be run from the system JOB$ table. New jobs selected are ordered by time.

2. The coordinator process dynamically spawns job queue slave processes (Jnnn) to run the jobs.

3. The job queue process runs one of the jobs that was selected by the CJQ0 process for execution. Each job queue process runs one job at a time to completion.

4. After the process finishes execution of a single job, it polls for more jobs. If no jobs are scheduled for execution, then it enters a sleep state, from which it wakes up at periodic intervals and polls for more jobs. If the process does not find any new jobs, then it terminates after a preset interval.

The initialization parameter JOB_QUEUE_PROCESSES represents the maximum number of job queue processes that can concurrently run on an instance. However, clients should not assume that all job queue processes are available for job execution.

The flashback data archive process (FBDA) archives historical rows of tracked tables into Flashback Data Archives. When a transaction containing DML on a tracked table commits, this process stores the pre-image of the changed rows into the Flashback Data Archive. It also keeps metadata on the current rows.

FBDA automatically manages the Flashback Data Archive for space, organization, and retention. Additionally, the process keeps track of how long the archiving of tracked transactions has occurred.

The SMCO process coordinates the execution of various space management related tasks, such as proactive space allocation and space reclamation. SMCO dynamically spawns slave processes (Wnnn) to implement the task.

I/O slave processes (Innn) simulate asynchronous I/O for systems and devices that do not support it. In asynchronous I/O, there is no timing requirement for transmission, enabling other processes to start before the transmission has finished.

For example, assume that an application writes 1000 blocks to a disk on an operating system that does not support asynchronous I/O. Each write occurs sequentially and waits for a confirmation that the write was successful. With asynchronous disk, the application can write the blocks in bulk and perform other work while waiting for a response from the operating system that all blocks were written.

To simulate asynchronous I/O, one process oversees several slave processes. The invoker process assigns work to each of the slave processes, who wait for each write to complete and report back to the invoker when done. In true asynchronous I/O the operating system waits for the I/O to complete and reports back to the process, while in simulated asynchronous I/O the slaves wait and report back to the invoker.

The database supports different types of I/O slaves, including the following:

• I/O slaves for Recovery Manager (RMAN)

  When using RMAN to back up or restore data, you can use I/O slaves for both disk and tape devices.

• Database writer slaves

  If it is not practical to use multiple database writer processes, such as when the computer has one CPU, then the database can distribute I/O over multiple slave processes. DBW is the only process that scans the buffer cache LRU list for blocks to be written to disk. However, I/O slaves perform the I/O for these blocks.

In parallel execution or parallel processing, multiple processes work together simultaneously to run a single SQL statement. By dividing the work among multiple processes, Oracle Database can run the statement more quickly. For example, four processes handle four different quarters in a year instead of one process handling all four quarters by itself.

Parallel execution contrasts with serial execution, in which a single server process performs all necessary processing for the sequential execution of a SQL statement. For example, to perform a full table scan such as SELECT * FROM employees, one server process performs all of the work, as shown in Figure 15-5.

Figure 15-5 Serial Full Table Scan

Description of "Figure 15-5 Serial Full Table Scan "

Parallel execution reduces response time for data-intensive operations on large databases such as data warehouses. Symmetric multiprocessing (SMP) and clustered system gain the largest performance benefits from parallel execution because statement processing can be split up among multiple CPUs. Parallel execution can also benefit certain types of OLTP and hybrid systems.

In Oracle RAC systems, the service placement of a specific service controls parallel execution. Specifically, parallel processes run on the nodes on which the service is configured. By default, Oracle Database runs parallel processes only on an instance that offers the service used to connect to the database. This does not affect other parallel operations such as parallel recovery or the processing of GV$ queries.