22.3.3.6 Defining NDB Cluster Data Nodes

NodeId

A unique node ID is used as the node's address for all cluster internal messages. For data nodes, this is an integer in the range 1 to 48 inclusive. Each node in the cluster must have a unique identifier.

NodeId is the only supported parameter name to use when identifying data nodes.

ExecuteOnComputer

This refers to the Id set for one of the computers defined in a [computer] section.

HostName

Specifying this parameter defines the hostname of the computer on which the data node is to reside. To specify a hostname other than localhost, either this parameter or ExecuteOnComputer is required.

ServerPort

Each node in the cluster uses a port to connect to other nodes. By default, this port is allocated dynamically in such a way as to ensure that no two nodes on the same host computer receive the same port number, so it should normally not be necessary to specify a value for this parameter.

However, if you need to be able to open specific ports in a firewall to permit communication between data nodes and API nodes (including SQL nodes), you can set this parameter to the number of the desired port in an [ndbd] section or (if you need to do this for multiple data nodes) the [ndbd default] section of the config.ini file, and then open the port having that number for incoming connections from SQL nodes, API nodes, or both.

TcpBind_INADDR_ANY

Setting this parameter to TRUE or 1 binds IP_ADDR_ANY so that connections can be made from anywhere (for autogenerated connections). The default is FALSE (0).

NodeGroup

This parameter can be used to assign a data node to a specific node group. It is read only when the cluster is started for the first time, and cannot be used to reassign a data node to a different node group online. It is generally not desirable to use this parameter in the [ndbd default] section of the config.ini file, and care must be taken not to assign nodes to node groups in such a way that an invalid numbers of nodes are assigned to any node groups.

The NodeGroup parameter is chiefly intended for use in adding a new node group to a running NDB Cluster without having to perform a rolling restart. For this purpose, you should set it to 65536 (the maximum value). You are not required to set a NodeGroup value for all cluster data nodes, only for those nodes which are to be started and added to the cluster as a new node group at a later time. For more information, see Section 22.5.15.3, “Adding NDB Cluster Data Nodes Online: Detailed Example”.

LocationDomainId

Assigns a data node to a specific availability domain (also known as an availability zone) within a cloud. By informing NDB which nodes are in which availability domains, performance can be improved in a cloud environment in the following ways:

LocationDomainId takes an integer value between 0 and 16 inclusive, with 0 being the default; using 0 is the same as leaving the parameter unset.

NoOfReplicas

This global parameter can be set only in the [ndbd default] section, and defines the number of replicas for each table stored in the cluster. This parameter also specifies the size of node groups. A node group is a set of nodes all storing the same information.

Node groups are formed implicitly. The first node group is formed by the set of data nodes with the lowest node IDs, the next node group by the set of the next lowest node identities, and so on. By way of example, assume that we have 4 data nodes and that NoOfReplicas is set to 2. The four data nodes have node IDs 2, 3, 4 and 5. Then the first node group is formed from nodes 2 and 3, and the second node group by nodes 4 and 5. It is important to configure the cluster in such a manner that nodes in the same node groups are not placed on the same computer because a single hardware failure would cause the entire cluster to fail.

If no node IDs are provided, the order of the data nodes will be the determining factor for the node group. Whether or not explicit assignments are made, they can be viewed in the output of the management client's SHOW command.

The default value for NoOfReplicas is 2. This is the recommended value for most production environments.

The value for this parameter must divide evenly into the number of data nodes in the cluster. For example, if there are two data nodes, then NoOfReplicas must be equal to either 1 or 2, since 2/3 and 2/4 both yield fractional values; if there are four data nodes, then NoOfReplicas must be equal to 1, 2, or 4.

DataDir

This parameter specifies the directory where trace files, log files, pid files and error logs are placed.

The default is the data node process working directory.

FileSystemPath

This parameter specifies the directory where all files created for metadata, REDO logs, UNDO logs (for Disk Data tables), and data files are placed. The default is the directory specified by DataDir.

The recommended directory hierarchy for NDB Cluster includes /var/lib/mysql-cluster, under which a directory for the node's file system is created. The name of this subdirectory contains the node ID. For example, if the node ID is 2, this subdirectory is named ndb_2_fs.

BackupDataDir

This parameter specifies the directory in which backups are placed.

DataMemory

This parameter defines the amount of space (in bytes) available for storing database records. The entire amount specified by this value is allocated in memory, so it is extremely important that the machine has sufficient physical memory to accommodate it.

The memory allocated by DataMemory is used to store both the actual records and indexes. There is a 16-byte overhead on each record; an additional amount for each record is incurred because it is stored in a 32KB page with 128 byte page overhead (see below). There is also a small amount wasted per page due to the fact that each record is stored in only one page.

For variable-size table attributes, the data is stored on separate data pages, allocated from DataMemory. Variable-length records use a fixed-size part with an extra overhead of 4 bytes to reference the variable-size part. The variable-size part has 2 bytes overhead plus 2 bytes per attribute.

The maximum record size is 14000 bytes.

Resources assigned to DataMemory are used for storing all data and indexes. (Any memory configured as IndexMemory is automatically added to that used by DataMemory to form a common resource pool.)

Currently, NDB Cluster can use a maximum of 512 MB for hash indexes per partition, which means in some cases it is possible to get Table is full errors in MySQL client applications even when ndb_mgm -e "ALL REPORT MEMORYUSAGE" shows significant free DataMemory. This can also pose a problem with data node restarts on nodes that are heavily loaded with data.

You can control the number of partitions per local data manager for a given table by setting the NDB_TABLE option PARTITION_BALANCE to one of the values FOR_RA_BY_LDM, FOR_RA_BY_LDM_X_2, FOR_RA_BY_LDM_X_3, or FOR_RA_BY_LDM_X_4, for 1, 2, 3, or 4 partitions per LDM, respectively, when creating the table (see Section 13.1.20.11, “Setting NDB_TABLE Options”).

You can also use the MinFreePct configuration parameter to help avoid problems with node restarts.

The memory space allocated by DataMemory consists of 32KB pages, which are allocated to table fragments. Each table is normally partitioned into the same number of fragments as there are data nodes in the cluster. Thus, for each node, there are the same number of fragments as are set in NoOfReplicas.

Once a page has been allocated, it is currently not possible to return it to the pool of free pages, except by deleting the table. (This also means that DataMemory pages, once allocated to a given table, cannot be used by other tables.) Performing a data node recovery also compresses the partition because all records are inserted into empty partitions from other live nodes.

The DataMemory memory space also contains UNDO information: For each update, a copy of the unaltered record is allocated in the DataMemory. There is also a reference to each copy in the ordered table indexes. Unique hash indexes are updated only when the unique index columns are updated, in which case a new entry in the index table is inserted and the old entry is deleted upon commit. For this reason, it is also necessary to allocate enough memory to handle the largest transactions performed by applications using the cluster. In any case, performing a few large transactions holds no advantage over using many smaller ones, for the following reasons:

The default value for DataMemory in NDB 8.0 is 98MB. The minimum value is 1MB. There is no maximum size, but in reality the maximum size has to be adapted so that the process does not start swapping when the limit is reached. This limit is determined by the amount of physical RAM available on the machine and by the amount of memory that the operating system may commit to any one process. 32-bit operating systems are generally limited to 2−4GB per process; 64-bit operating systems can use more. For large databases, it may be preferable to use a 64-bit operating system for this reason.

IndexMemory

The IndexMemory parameter is deprecated (and subject to future removal); any memory assigned to IndexMemory is allocated instead to the same pool as DataMemory, which is solely responsible for all resources needed for storing data and indexes in memory. In NDB 8.0, the use of IndexMemory in the cluster configuration file triggers a warning from the management server.

You can estimate the size of a hash index using this formula:

  size  = ( (fragments * 32K) + (rows * 18) )
          * replicas

fragments is the number of fragments, replicas is the number of replicas (normally 2), and rows is the number of rows. If a table has one million rows, 8 fragments, and 2 replicas, the expected index memory usage is calculated as shown here:

  ((8 * 32K) + (1000000 * 18)) * 2 = ((8 * 32768) + (1000000 * 18)) * 2
  = (262144 + 18000000) * 2
  = 18262144 * 2 = 36524288 bytes = ~35MB

Index statistics for ordered indexes (when these are enabled) are stored in the mysql.ndb_index_stat_sample table. Since this table has a hash index, this adds to index memory usage. An upper bound to the number of rows for a given ordered index can be calculated as follows:

  sample_size= key_size + ((key_attributes + 1) * 4)

  sample_rows = IndexStatSaveSize
                * ((0.01 * IndexStatSaveScale * log2(rows * sample_size)) + 1)
                / sample_size

In the preceding formula, key_size is the size of the ordered index key in bytes, key_attributes is the number ot attributes in the ordered index key, and rows is the number of rows in the base table.

Assume that table t1 has 1 million rows and an ordered index named ix1 on two four-byte integers. Assume in addition that IndexStatSaveSize and IndexStatSaveScale are set to their default values (32K and 100, respectively). Using the previous 2 formulas, we can calculate as follows:

  sample_size = 8  + ((1 + 2) * 4) = 20 bytes

  sample_rows = 32K
                * ((0.01 * 100 * log2(1000000*20)) + 1)
                / 20
                = 32768 * ( (1 * ~16.811) +1) / 20
                = 32768 * ~17.811 / 20
                = ~29182 rows

The expected index memory usage is thus 2 * 18 * 29182 = ~1050550 bytes.

In NDB 8.0, the minimum and default vaue for this parameter is 0 (zero).

StringMemory

This parameter determines how much memory is allocated for strings such as table names, and is specified in an [ndbd] or [ndbd default] section of the config.ini file. A value between 0 and 100 inclusive is interpreted as a percent of the maximum default value, which is calculated based on a number of factors including the number of tables, maximum table name size, maximum size of .FRM files, MaxNoOfTriggers, maximum column name size, and maximum default column value.

A value greater than 100 is interpreted as a number of bytes.

The default value is 25—that is, 25 percent of the default maximum.

Under most circumstances, the default value should be sufficient, but when you have a great many NDB tables (1000 or more), it is possible to get Error 773 Out of string memory, please modify StringMemory config parameter: Permanent error: Schema error, in which case you should increase this value. 25 (25 percent) is not excessive, and should prevent this error from recurring in all but the most extreme conditions.

MaxNoOfConcurrentTransactions

Each cluster data node requires a transaction record for each active transaction in the cluster. The task of coordinating transactions is distributed among all of the data nodes. The total number of transaction records in the cluster is the number of transactions in any given node times the number of nodes in the cluster.

Transaction records are allocated to individual MySQL servers. Each connection to a MySQL server requires at least one transaction record, plus an additional transaction object per table accessed by that connection. This means that a reasonable minimum for the total number of transactions in the cluster can be expressed as

TotalNoOfConcurrentTransactions =
    (maximum number of tables accessed in any single transaction + 1)
    * number of SQL nodes

Suppose that there are 10 SQL nodes using the cluster. A single join involving 10 tables requires 11 transaction records; if there are 10 such joins in a transaction, then 10 * 11 = 110 transaction records are required for this transaction, per MySQL server, or 110 * 10 = 1100 transaction records total. Each data node can be expected to handle TotalNoOfConcurrentTransactions / number of data nodes. For an NDB Cluster having 4 data nodes, this would mean setting MaxNoOfConcurrentTransactions on each data node to 1100 / 4 = 275. In addition, you should provide for failure recovery by ensuring that a single node group can accommodate all concurrent transactions; in other words, that each data node's MaxNoOfConcurrentTransactions is sufficient to cover a number of transactions equal to TotalNoOfConcurrentTransactions / number of node groups. If this cluster has a single node group, then MaxNoOfConcurrentTransactions should be set to 1100 (the same as the total number of concurrent transactions for the entire cluster).

In addition, each transaction involves at least one operation; for this reason, the value set for MaxNoOfConcurrentTransactions should always be no more than the value of MaxNoOfConcurrentOperations.

This parameter must be set to the same value for all cluster data nodes. This is due to the fact that, when a data node fails, the oldest surviving node re-creates the transaction state of all transactions that were ongoing in the failed node.

It is possible to change this value using a rolling restart, but the amount of traffic on the cluster must be such that no more transactions occur than the lower of the old and new levels while this is taking place.

The default value is 4096.

MaxNoOfConcurrentOperations

It is a good idea to adjust the value of this parameter according to the size and number of transactions. When performing transactions which involve only a few operations and records, the default value for this parameter is usually sufficient. Performing large transactions involving many records usually requires that you increase its value.

Records are kept for each transaction updating cluster data, both in the transaction coordinator and in the nodes where the actual updates are performed. These records contain state information needed to find UNDO records for rollback, lock queues, and other purposes.

This parameter should be set at a minimum to the number of records to be updated simultaneously in transactions, divided by the number of cluster data nodes. For example, in a cluster which has four data nodes and which is expected to handle one million concurrent updates using transactions, you should set this value to 1000000 / 4 = 250000. To help provide resiliency against failures, it is suggested that you set this parameter to a value that is high enough to permit an individual data node to handle the load for its node group. In other words, you should set the value equal to total number of concurrent operations / number of node groups. (In the case where there is a single node group, this is the same as the total number of concurrent operations for the entire cluster.)

Because each transaction always involves at least one operation, the value of MaxNoOfConcurrentOperations should always be greater than or equal to the value of MaxNoOfConcurrentTransactions.

Read queries which set locks also cause operation records to be created. Some extra space is allocated within individual nodes to accommodate cases where the distribution is not perfect over the nodes.

When queries make use of the unique hash index, there are actually two operation records used per record in the transaction. The first record represents the read in the index table and the second handles the operation on the base table.

The default value is 32768.

This parameter actually handles two values that can be configured separately. The first of these specifies how many operation records are to be placed with the transaction coordinator. The second part specifies how many operation records are to be local to the database.

A very large transaction performed on an eight-node cluster requires as many operation records in the transaction coordinator as there are reads, updates, and deletes involved in the transaction. However, the operation records of the are spread over all eight nodes. Thus, if it is necessary to configure the system for one very large transaction, it is a good idea to configure the two parts separately. MaxNoOfConcurrentOperations will always be used to calculate the number of operation records in the transaction coordinator portion of the node.

It is also important to have an idea of the memory requirements for operation records. These consume about 1KB per record.

MaxNoOfLocalOperations

By default, this parameter is calculated as 1.1 × MaxNoOfConcurrentOperations. This fits systems with many simultaneous transactions, none of them being very large. If there is a need to handle one very large transaction at a time and there are many nodes, it is a good idea to override the default value by explicitly specifying this parameter.

MaxDMLOperationsPerTransaction

This parameter limits the size of a transaction. The transaction is aborted if it requires more than this many DML operations. The minimum number of operations per transaction is 32; however, you can set MaxDMLOperationsPerTransaction to 0 to disable any limitation on the number of DML operations per transaction. The maximum (and default) is 4294967295.

The value of this parameter cannot exceed that set for MaxNoOfConcurrentOperations.

MaxNoOfConcurrentIndexOperations

For queries using a unique hash index, another temporary set of operation records is used during a query's execution phase. This parameter sets the size of that pool of records. Thus, this record is allocated only while executing a part of a query. As soon as this part has been executed, the record is released. The state needed to handle aborts and commits is handled by the normal operation records, where the pool size is set by the parameter MaxNoOfConcurrentOperations.

The default value of this parameter is 8192. Only in rare cases of extremely high parallelism using unique hash indexes should it be necessary to increase this value. Using a smaller value is possible and can save memory if the DBA is certain that a high degree of parallelism is not required for the cluster.

MaxNoOfFiredTriggers

The default value of MaxNoOfFiredTriggers is 4000, which is sufficient for most situations. In some cases it can even be decreased if the DBA feels certain the need for parallelism in the cluster is not high.

A record is created when an operation is performed that affects a unique hash index. Inserting or deleting a record in a table with unique hash indexes or updating a column that is part of a unique hash index fires an insert or a delete in the index table. The resulting record is used to represent this index table operation while waiting for the original operation that fired it to complete. This operation is short-lived but can still require a large number of records in its pool for situations with many parallel write operations on a base table containing a set of unique hash indexes.

TransactionBufferMemory

The memory affected by this parameter is used for tracking operations fired when updating index tables and reading unique indexes. This memory is used to store the key and column information for these operations. It is only very rarely that the value for this parameter needs to be altered from the default.

The default value for TransactionBufferMemory is 1MB.

Normal read and write operations use a similar buffer, whose usage is even more short-lived. The compile-time parameter ZATTRBUF_FILESIZE (found in ndb/src/kernel/blocks/Dbtc/Dbtc.hpp) set to 4000 × 128 bytes (500KB). A similar buffer for key information, ZDATABUF_FILESIZE (also in Dbtc.hpp) contains 4000 × 16 = 62.5KB of buffer space. Dbtc is the module that handles transaction coordination.

ReservedConcurrentIndexOperations

Number of simultaneous index operations having dedicated resources on one data node.

ReservedConcurrentOperations

Number of simultaneous operations having dedicated resources in transaction coordinators on one data node.

ReservedConcurrentScans

Number of simultaneous scans having dedicated resources on one data node.

ReservedConcurrentTransactions

Number of simultaneous transactions having dedicated resources on one data node.

ReservedFiredTriggers

Number of triggers that have dedicated resources on one ndbd(DB) node.

ReservedLocalScans

Number of simultaneous fragment scans having dedicated resources on one data node.

ReservedTransactionBufferMemory

Dynamic buffer space (in bytes) for key and attribute data allocated to each data node.

BatchSizePerLocalScan

This parameter is used to calculate the number of lock records used to handle concurrent scan operations.

BatchSizePerLocalScan has a strong connection to the BatchSize defined in the SQL nodes.

LongMessageBuffer

This is an internal buffer used for passing messages within individual nodes and between nodes. The default is 64MB.

This parameter seldom needs to be changed from the default.

MaxFKBuildBatchSize

Maximum scan batch size used for building foreign keys. Increasing the value set for this parameter may speed up building of foreign key builds at the expense of greater impact to ongoing traffic.

MaxNoOfConcurrentScans

This parameter is used to control the number of parallel scans that can be performed in the cluster. Each transaction coordinator can handle the number of parallel scans defined for this parameter. Each scan query is performed by scanning all partitions in parallel. Each partition scan uses a scan record in the node where the partition is located, the number of records being the value of this parameter times the number of nodes. The cluster should be able to sustain MaxNoOfConcurrentScans scans concurrently from all nodes in the cluster.

Scans are actually performed in two cases. The first of these cases occurs when no hash or ordered indexes exists to handle the query, in which case the query is executed by performing a full table scan. The second case is encountered when there is no hash index to support the query but there is an ordered index. Using the ordered index means executing a parallel range scan. The order is kept on the local partitions only, so it is necessary to perform the index scan on all partitions.

The default value of MaxNoOfConcurrentScans is 256. The maximum value is 500.

MaxNoOfLocalScans

Specifies the number of local scan records if many scans are not fully parallelized. When the number of local scan records is not provided, it is calculated as shown here:

4 * MaxNoOfConcurrentScans * [# data nodes] + 2

The minimum value is 32.

MaxParallelCopyInstances

This parameter sets the parallelization used in the copy phase of a node restart or system restart, when a node that is currently just starting is synchronised with a node that already has current data by copying over any changed records from the node that is up to date. Because full parallelism in such cases can lead to overload situations, MaxParallelCopyInstances provides a means to decrease it. This parameter's default value 0. This value means that the effective parallelism is equal to the number of LDM instances in the node just starting as well as the node updating it.

MaxParallelScansPerFragment

It is possible to configure the maximum number of parallel scans (TUP scans and TUX scans) allowed before they begin queuing for serial handling. You can increase this to take advantage of any unused CPU when performing large number of scans in parallel and improve their performance.

The default value for this parameter is 256.

MaxReorgBuildBatchSize

Maximum scan batch size used for reorganization of table partitions. Increasing the value set for this parameter may speed up reorganization at the expense of greater impact to ongoing traffic.

MaxUIBuildBatchSize

Maximum scan batch size used for building unique keys. Increasing the value set for this parameter may speed up such builds at the expense of greater impact to ongoing traffic.

FragmentLogFileSize

Setting this parameter enables you to control directly the size of redo log files. This can be useful in situations when NDB Cluster is operating under a high load and it is unable to close fragment log files quickly enough before attempting to open new ones (only 2 fragment log files can be open at one time); increasing the size of the fragment log files gives the cluster more time before having to open each new fragment log file. The default value for this parameter is 16M.

For more information about fragment log files, see the description for NoOfFragmentLogFiles.

InitialNoOfOpenFiles

This parameter sets the initial number of internal threads to allocate for open files.

The default value is 27.

InitFragmentLogFiles

By default, fragment log files are created sparsely when performing an initial start of a data node—that is, depending on the operating system and file system in use, not all bytes are necessarily written to disk. However, it is possible to override this behavior and force all bytes to be written, regardless of the platform and file system type being used, by means of this parameter. InitFragmentLogFiles takes either of two values:

Depending on your operating system and file system, setting InitFragmentLogFiles=FULL may help eliminate I/O errors on writes to the REDO log.

EnablePartialLcp

When true, enable partial local checkpoints: This means that each LCP records only part of the full database, plus any records containing rows changed since the last LCP; if no rows have changed, the LCP updates only the LCP control file and does not update any data files.

If EnablePartialLcp is disabled (false), each LCP uses only a single file and writes a full checkpoint; this requires the least amount of disk space for LCPs, but increases the write load for each LCP. The default value is enabled (true). The proportion of space used by partiaL LCPS can be modified by the setting for the RecoveryWork configuration parameter.

Setting this parameter to false also disables the calculation of disk write speed used by the adaptive LCP control mechanism.

LcpScanProgressTimeout

A local checkpoint fragment scan watchdog checks periodically for no progress in each fragment scan performed as part of a local checkpoint, and shuts down the node if there is no progress after a given amount of time has elapsed. This interval can be set using the LcpScanProgressTimeout data node configuration parameter, which sets the maximum time for which the local checkpoint can be stalled before the LCP fragment scan watchdog shuts down the node.

The default value is 60 seconds (providing compatibility with previous releases). Setting this parameter to 0 disables the LCP fragment scan watchdog altogether.

MaxNoOfOpenFiles

This parameter sets a ceiling on how many internal threads to allocate for open files. Any situation requiring a change in this parameter should be reported as a bug.

The default value is 0. However, the minimum value to which this parameter can be set is 20.

MaxNoOfSavedMessages

This parameter sets the maximum number of errors written in the error log as well as the maximum number of trace files that are kept before overwriting the existing ones. Trace files are generated when, for whatever reason, the node crashes.

The default is 25, which sets these maximums to 25 error messages and 25 trace files.

MaxLCPStartDelay

In parallel data node recovery, only table data is actually copied and synchronized in parallel; synchronization of metadata such as dictionary and checkpoint information is done in a serial fashion. In addition, recovery of dictionary and checkpoint information cannot be executed in parallel with performing of local checkpoints. This means that, when starting or restarting many data nodes concurrently, data nodes may be forced to wait while a local checkpoint is performed, which can result in longer node recovery times.

It is possible to force a delay in the local checkpoint to permit more (and possibly all) data nodes to complete metadata synchronization; once each data node's metadata synchronization is complete, all of the data nodes can recover table data in parallel, even while the local checkpoint is being executed. To force such a delay, set MaxLCPStartDelay, which determines the number of seconds the cluster can wait to begin a local checkpoint while data nodes continue to synchronize metadata. This parameter should be set in the [ndbd default] section of the config.ini file, so that it is the same for all data nodes. The maximum value is 600; the default is 0.

NoOfFragmentLogFiles

This parameter sets the number of REDO log files for the node, and thus the amount of space allocated to REDO logging. Because the REDO log files are organized in a ring, it is extremely important that the first and last log files in the set (sometimes referred to as the “head” and “tail” log files, respectively) do not meet. When these approach one another too closely, the node begins aborting all transactions encompassing updates due to a lack of room for new log records.

A REDO log record is not removed until both required local checkpoints have been completed since that log record was inserted. Checkpointing frequency is determined by its own set of configuration parameters discussed elsewhere in this chapter.

The default parameter value is 16, which by default means 16 sets of 4 16MB files for a total of 1024MB. The size of the individual log files is configurable using the FragmentLogFileSize parameter. In scenarios requiring a great many updates, the value for NoOfFragmentLogFiles may need to be set as high as 300 or even higher to provide sufficient space for REDO logs.

If the checkpointing is slow and there are so many writes to the database that the log files are full and the log tail cannot be cut without jeopardizing recovery, all updating transactions are aborted with internal error code 410 (Out of log file space temporarily). This condition prevails until a checkpoint has completed and the log tail can be moved forward.

RecoveryWork

Percentage of storage overhead for LCP files. This parameter has an effect only when EnablePartialLcp is true, that is, only when partial local checkpoints are enabled. A higher value means:

A lower value for RecoveryWork means:

For example, setting RecoveryWork to 60 means that the total size of an LCP is roughly 1 + 0.6 = 1.6 times the size of the data to be checkpointed. This means that 60% more work is required during the restore phase of a restart compared to the work done during a restart that uses full checkpoints. (This is more than compensated for during other phases of the restart such that the restart as a whole is still faster when using partial LCPs than when using full LCPs.) In order not to fill up the redo log, it is necessary to write at 1 + (1 / RecoveryWork) times the rate of data changes during checkpoints—thus, when RecoveryWork = 60, it is necessary to write at approximately 1 + (1 / 0.6 ) = 2.67 times the change rate. In other words, if changes are being written at 10 MByte per second, the checkpoint needs to be written at roughly 26.7 MByte per second.

Setting RecoveryWork = 40 means that only 1.4 times the total LCP size is needed (and thus the restore phase takes 10 to 15 percent less time. In this case, the checkpoint write rate is 3.5 times the rate of change.

The NDB source distribution includes a test program for simulating LCPs. lcp_simulator.cc can be found in storage/ndb/src/kernel/blocks/backup/. To compile and run it on Unix platforms, execute the commands shown here:

shell> gcc lcp_simulator.cc
shell> ./a.out

This program has no dependencies other than stdio.h, and does not require a connection to an NDB cluster or a MySQL server. By default, it simulates 300 LCPs (three sets of 100 LCPs, each consisting of inserts, updates, and deletes, in turn), reporting the size of the LCP after each one. You can alter the simulation by changing the values of recovery_work, insert_work, and delete_work in the source and recompiling. For more information, see the source of the program.

InsertRecoveryWork

Percentage of RecoveryWork used for inserted rows. A higher value increases the number of writes during a local checkpoint, and decreases the total size of the LCP. A lower value decreases the number of writes during an LCP, but results in more space being used for the LCP, which means that recovery takes longer. This parameter has an effect only when EnablePartialLcp is true, that is, only when partial local checkpoints are enabled.

EnableRedoControl

Enable adaptive checkpointing speed for controlling redo log usage. Set to false to disable (the default). Setting EnablePartialLcp to false also disables the adaptive calculation.

When enabled, EnableRedoControl allows the data nodes greater flexibility with regard to the rate at which they write LCPs to disk. More specifically, enabling this parameter means that higher write rates can be employed, so that LCPs can complete and Redo logs be trimmed more quickly, thereby reducing recovery time and disk space requirements. This functionality allows data nodes to make better use of the higher rate of I/O and greater bandwidth available from modern solid-state storage devices and protocols, such as solid-state drives (SSDs) using Non-Volatile Memory Express (NVMe).

The parameter currently defaults to false (disabled) due to the fact that NDB is still deployed widely on systems whose I/O or bandwidth is constrained relative to those employing solid-state technology, such as those using conventional hard disks (HDDs). In settings such as these, the EnableRedoControl mechanism can easily cause the I/O subsystem to become saturated, increasing wait times for data node input and output. In particular, this can cause issues with NDB Disk Data tables which have tablespaces or log file groups sharing a constrained IO subsystem with data node LCP and redo log files; such problems potentially include node or cluster failure due to GCP stop errors.

MaxNoOfAttributes

This parameter sets a suggested maximum number of attributes that can be defined in the cluster; like MaxNoOfTables, it is not intended to function as a hard upper limit.

(In older NDB Cluster releases, this parameter was sometimes treated as a hard limit for certain operations. This caused problems with NDB Cluster Replication, when it was possible to create more tables than could be replicated, and sometimes led to confusion when it was possible [or not possible, depending on the circumstances] to create more than MaxNoOfAttributes attributes.)

The default value is 1000, with the minimum possible value being 32. The maximum is 4294967039. Each attribute consumes around 200 bytes of storage per node due to the fact that all metadata is fully replicated on the servers.

When setting MaxNoOfAttributes, it is important to prepare in advance for any ALTER TABLE statements that you might want to perform in the future. This is due to the fact, during the execution of ALTER TABLE on a Cluster table, 3 times the number of attributes as in the original table are used, and a good practice is to permit double this amount. For example, if the NDB Cluster table having the greatest number of attributes (greatest_number_of_attributes) has 100 attributes, a good starting point for the value of MaxNoOfAttributes would be 6 * greatest_number_of_attributes = 600.

You should also estimate the average number of attributes per table and multiply this by MaxNoOfTables. If this value is larger than the value obtained in the previous paragraph, you should use the larger value instead.

Assuming that you can create all desired tables without any problems, you should also verify that this number is sufficient by trying an actual ALTER TABLE after configuring the parameter. If this is not successful, increase MaxNoOfAttributes by another multiple of MaxNoOfTables and test it again.

MaxNoOfTables

A table object is allocated for each table and for each unique hash index in the cluster. This parameter sets a suggested maximum number of table objects for the cluster as a whole; like MaxNoOfAttributes, it is not intended to function as a hard upper limit.

(In older NDB Cluster releases, this parameter was sometimes treated as a hard limit for certain operations. This caused problems with NDB Cluster Replication, when it was possible to create more tables than could be replicated, and sometimes led to confusion when it was possible [or not possible, depending on the circumstances] to create more than MaxNoOfTables tables.)

For each attribute that has a BLOB data type an extra table is used to store most of the BLOB data. These tables also must be taken into account when defining the total number of tables.

The default value of this parameter is 128. The minimum is 8 and the maximum is 20320. Each table object consumes approximately 20KB per node.

MaxNoOfOrderedIndexes

For each ordered index in the cluster, an object is allocated describing what is being indexed and its storage segments. By default, each index so defined also defines an ordered index. Each unique index and primary key has both an ordered index and a hash index. MaxNoOfOrderedIndexes sets the total number of ordered indexes that can be in use in the system at any one time.

The default value of this parameter is 128. Each index object consumes approximately 10KB of data per node.

MaxNoOfUniqueHashIndexes

For each unique index that is not a primary key, a special table is allocated that maps the unique key to the primary key of the indexed table. By default, an ordered index is also defined for each unique index. To prevent this, you must specify the USING HASH option when defining the unique index.

The default value is 64. Each index consumes approximately 15KB per node.

MaxNoOfTriggers

Internal update, insert, and delete triggers are allocated for each unique hash index. (This means that three triggers are created for each unique hash index.) However, an ordered index requires only a single trigger object. Backups also use three trigger objects for each normal table in the cluster.

Replication between clusters also makes use of internal triggers.

This parameter sets the maximum number of trigger objects in the cluster.

The default value is 768.

MaxNoOfSubscriptions

Each NDB table in an NDB Cluster requires a subscription in the NDB kernel. For some NDB API applications, it may be necessary or desirable to change this parameter. However, for normal usage with MySQL servers acting as SQL nodes, there is not any need to do so.

The default value for MaxNoOfSubscriptions is 0, which is treated as equal to MaxNoOfTables. Each subscription consumes 108 bytes.

MaxNoOfSubscribers

This parameter is of interest only when using NDB Cluster Replication. The default value is 0, which is treated as 2 * MaxNoOfTables; that is, there is one subscription per NDB table for each of two MySQL servers (one acting as the replication master and the other as the slave). Each subscriber uses 16 bytes of memory.

When using circular replication, multi-master replication, and other replication setups involving more than 2 MySQL servers, you should increase this parameter to the number of mysqld processes included in replication (this is often, but not always, the same as the number of clusters). For example, if you have a circular replication setup using three NDB Cluster s, with one mysqld attached to each cluster, and each of these mysqld processes acts as a master and as a slave, you should set MaxNoOfSubscribers equal to 3 * MaxNoOfTables.

For more information, see Section 22.6, “NDB Cluster Replication”.

MaxNoOfConcurrentSubOperations

This parameter sets a ceiling on the number of operations that can be performed by all API nodes in the cluster at one time. The default value (256) is sufficient for normal operations, and might need to be adjusted only in scenarios where there are a great many API nodes each performing a high volume of operations concurrently.

CompressedBackup

Enabling this parameter causes backup files to be compressed. The compression used is equivalent to gzip --fast, and can save 50% or more of the space required on the data node to store uncompressed backup files. Compressed backups can be enabled for individual data nodes, or for all data nodes (by setting this parameter in the [ndbd default] section of the config.ini file).

The default value is 0 (disabled).

CompressedLCP

Setting this parameter to 1 causes local checkpoint files to be compressed. The compression used is equivalent to gzip --fast, and can save 50% or more of the space required on the data node to store uncompressed checkpoint files. Compressed LCPs can be enabled for individual data nodes, or for all data nodes (by setting this parameter in the [ndbd default] section of the config.ini file).

The default value is 0 (disabled).

CrashOnCorruptedTuple

When this parameter is enabled (the default), it forces a data node to shut down whenever it encounters a corrupted tuple.

Diskless

It is possible to specify NDB Cluster tables as diskless, meaning that tables are not checkpointed to disk and that no logging occurs. Such tables exist only in main memory. A consequence of using diskless tables is that neither the tables nor the records in those tables survive a crash. However, when operating in diskless mode, it is possible to run ndbd on a diskless computer.

When this feature is enabled, Cluster online backup is disabled. In addition, a partial start of the cluster is not possible.

Diskless is disabled by default.

LateAlloc

Allocate memory for this data node after a connection to the management server has been established. Enabled by default.

LockPagesInMainMemory

For a number of operating systems, including Solaris and Linux, it is possible to lock a process into memory and so avoid any swapping to disk. This can be used to help guarantee the cluster's real-time characteristics.

This parameter takes one of the integer values 0, 1, or 2, which act as shown in the following list:

If the operating system is not configured to permit unprivileged users to lock pages, then the data node process making use of this parameter may have to be run as system root. (LockPagesInMainMemory uses the mlockall function. From Linux kernel 2.6.9, unprivileged users can lock memory as limited by max locked memory. For more information, see ulimit -l and http://linux.die.net/man/2/mlock).

ODirect

Enabling this parameter causes NDB to attempt using O_DIRECT writes for LCP, backups, and redo logs, often lowering kswapd and CPU usage. When using NDB Cluster on Linux, enable ODirect if you are using a 2.6 or later kernel.

ODirect is disabled by default.

ODirectSyncFlag

When this parameter is enabled, redo log writes are performed such that each completed file system write is handled as a call to fsync. The setting for this parameter is ignored if at least one of the following conditions is true:

Disabled by default.

RestartOnErrorInsert

This feature is accessible only when building the debug version where it is possible to insert errors in the execution of individual blocks of code as part of testing.

This feature is disabled by default.

StopOnError

This parameter specifies whether a data node process should exit or perform an automatic restart when an error condition is encountered.

This parameter's default value is 1; this means that, by default, an error causes the data node process to halt.

When an error is encountered and StopOnError is 0, the data node process is restarted.

Users of MySQL Cluster Manager should note that, when StopOnError equals 1, this prevents the MySQL Cluster Manager agent from restarting any data nodes after it has performed its own restart and recovery. See Starting and Stopping the Agent on Linux, for more information.

UseShm

Enable a shared memory connection between this data node and the API node also running on this host. Set to 1 to enable.

TimeBetweenWatchDogCheck

To prevent the main thread from getting stuck in an endless loop at some point, a “watchdog” thread checks the main thread. This parameter specifies the number of milliseconds between checks. If the process remains in the same state after three checks, the watchdog thread terminates it.

This parameter can easily be changed for purposes of experimentation or to adapt to local conditions. It can be specified on a per-node basis although there seems to be little reason for doing so.

The default timeout is 6000 milliseconds (6 seconds).

TimeBetweenWatchDogCheckInitial

This is similar to the TimeBetweenWatchDogCheck parameter, except that TimeBetweenWatchDogCheckInitial controls the amount of time that passes between execution checks inside a storage node in the early start phases during which memory is allocated.

The default timeout is 6000 milliseconds (6 seconds).

StartPartialTimeout

This parameter specifies how long the Cluster waits for all data nodes to come up before the cluster initialization routine is invoked. This timeout is used to avoid a partial Cluster startup whenever possible.

This parameter is overridden when performing an initial start or initial restart of the cluster.

The default value is 30000 milliseconds (30 seconds). 0 disables the timeout, in which case the cluster may start only if all nodes are available.

StartPartitionedTimeout

If the cluster is ready to start after waiting for StartPartialTimeout milliseconds but is still possibly in a partitioned state, the cluster waits until this timeout has also passed. If StartPartitionedTimeout is set to 0, the cluster waits indefinitely.

This parameter is overridden when performing an initial start or initial restart of the cluster.

The default timeout is 60000 milliseconds (60 seconds).

StartFailureTimeout

If a data node has not completed its startup sequence within the time specified by this parameter, the node startup fails. Setting this parameter to 0 (the default value) means that no data node timeout is applied.

For nonzero values, this parameter is measured in milliseconds. For data nodes containing extremely large amounts of data, this parameter should be increased. For example, in the case of a data node containing several gigabytes of data, a period as long as 10−15 minutes (that is, 600000 to 1000000 milliseconds) might be required to perform a node restart.

StartNoNodeGroupTimeout

When a data node is configured with Nodegroup = 65536, is regarded as not being assigned to any node group. When that is done, the cluster waits StartNoNodegroupTimeout milliseconds, then treats such nodes as though they had been added to the list passed to the --nowait-nodes option, and starts. The default value is 15000 (that is, the management server waits 15 seconds). Setting this parameter equal to 0 means that the cluster waits indefinitely.

StartNoNodegroupTimeout must be the same for all data nodes in the cluster; for this reason, you should always set it in the [ndbd default] section of the config.ini file, rather than for individual data nodes.

See Section 22.5.15, “Adding NDB Cluster Data Nodes Online”, for more information.

HeartbeatIntervalDbDb

One of the primary methods of discovering failed nodes is by the use of heartbeats. This parameter states how often heartbeat signals are sent and how often to expect to receive them. Heartbeats cannot be disabled.

After missing four heartbeat intervals in a row, the node is declared dead. Thus, the maximum time for discovering a failure through the heartbeat mechanism is five times the heartbeat interval.

The default heartbeat interval is 5000 milliseconds (5 seconds). This parameter must not be changed drastically and should not vary widely between nodes. If one node uses 5000 milliseconds and the node watching it uses 1000 milliseconds, obviously the node will be declared dead very quickly. This parameter can be changed during an online software upgrade, but only in small increments.

See also Network communication and latency, as well as the description of the ConnectCheckIntervalDelay configuration parameter.

HeartbeatIntervalDbApi

Each data node sends heartbeat signals to each MySQL server (SQL node) to ensure that it remains in contact. If a MySQL server fails to send a heartbeat in time it is declared “dead,” in which case all ongoing transactions are completed and all resources released. The SQL node cannot reconnect until all activities initiated by the previous MySQL instance have been completed. The three-heartbeat criteria for this determination are the same as described for HeartbeatIntervalDbDb.

The default interval is 1500 milliseconds (1.5 seconds). This interval can vary between individual data nodes because each data node watches the MySQL servers connected to it, independently of all other data nodes.

For more information, see Network communication and latency.

HeartbeatOrder

Data nodes send heartbeats to one another in a circular fashion whereby each data node monitors the previous one. If a heartbeat is not detected by a given data node, this node declares the previous data node in the circle “dead” (that is, no longer accessible by the cluster). The determination that a data node is dead is done globally; in other words; once a data node is declared dead, it is regarded as such by all nodes in the cluster.

It is possible for heartbeats between data nodes residing on different hosts to be too slow compared to heartbeats between other pairs of nodes (for example, due to a very low heartbeat interval or temporary connection problem), such that a data node is declared dead, even though the node can still function as part of the cluster. .

In this type of situation, it may be that the order in which heartbeats are transmitted between data nodes makes a difference as to whether or not a particular data node is declared dead. If this declaration occurs unnecessarily, this can in turn lead to the unnecessary loss of a node group and as thus to a failure of the cluster.

Consider a setup where there are 4 data nodes A, B, C, and D running on 2 host computers host1 and host2, and that these data nodes make up 2 node groups, as shown in the following table:

Suppose the heartbeats are transmitted in the order A->B->C->D->A. In this case, the loss of the heartbeat between the hosts causes node B to declare node A dead and node C to declare node B dead. This results in loss of Node Group 0, and so the cluster fails. On the other hand, if the order of transmission is A->B->D->C->A (and all other conditions remain as previously stated), the loss of the heartbeat causes nodes A and D to be declared dead; in this case, each node group has one surviving node, and the cluster survives.

The HeartbeatOrder configuration parameter makes the order of heartbeat transmission user-configurable. The default value for HeartbeatOrder is zero; allowing the default value to be used on all data nodes causes the order of heartbeat transmission to be determined by NDB. If this parameter is used, it must be set to a nonzero value (maximum 65535) for every data node in the cluster, and this value must be unique for each data node; this causes the heartbeat transmission to proceed from data node to data node in the order of their HeartbeatOrder values from lowest to highest (and then directly from the data node having the highest HeartbeatOrder to the data node having the lowest value, to complete the circle). The values need not be consecutive. For example, to force the heartbeat transmission order A->B->D->C->A in the scenario outlined previously, you could set the HeartbeatOrder values as shown here:

To use this parameter to change the heartbeat transmission order in a running NDB Cluster, you must first set HeartbeatOrder for each data node in the cluster in the global configuration (config.ini) file (or files). To cause the change to take effect, you must perform either of the following:

You can use DUMP 908 to observe the effect of this parameter in the data node logs.

ConnectCheckIntervalDelay

This parameter enables connection checking between data nodes after one of them has failed heartbeat checks for 5 intervals of up to HeartbeatIntervalDbDb milliseconds.

Such a data node that further fails to respond within an interval of ConnectCheckIntervalDelay milliseconds is considered suspect, and is considered dead after two such intervals. This can be useful in setups with known latency issues.

The default value for this parameter is 0 (disabled).

TimeBetweenLocalCheckpoints

This parameter is an exception in that it does not specify a time to wait before starting a new local checkpoint; rather, it is used to ensure that local checkpoints are not performed in a cluster where relatively few updates are taking place. In most clusters with high update rates, it is likely that a new local checkpoint is started immediately after the previous one has been completed.

The size of all write operations executed since the start of the previous local checkpoints is added. This parameter is also exceptional in that it is specified as the base-2 logarithm of the number of 4-byte words, so that the default value 20 means 4MB (4 × 2²⁰) of write operations, 21 would mean 8MB, and so on up to a maximum value of 31, which equates to 8GB of write operations.

All the write operations in the cluster are added together. Setting TimeBetweenLocalCheckpoints to 6 or less means that local checkpoints will be executed continuously without pause, independent of the cluster's workload.

TimeBetweenGlobalCheckpoints

When a transaction is committed, it is committed in main memory in all nodes on which the data is mirrored. However, transaction log records are not flushed to disk as part of the commit. The reasoning behind this behavior is that having the transaction safely committed on at least two autonomous host machines should meet reasonable standards for durability.

It is also important to ensure that even the worst of cases—a complete crash of the cluster—is handled properly. To guarantee that this happens, all transactions taking place within a given interval are put into a global checkpoint, which can be thought of as a set of committed transactions that has been flushed to disk. In other words, as part of the commit process, a transaction is placed in a global checkpoint group. Later, this group's log records are flushed to disk, and then the entire group of transactions is safely committed to disk on all computers in the cluster.

This parameter defines the interval between global checkpoints. The default is 2000 milliseconds.

TimeBetweenGlobalCheckpointsTimeout

This parameter defines the minimum timeout between global checkpoints. The default is 120000 milliseconds.

TimeBetweenEpochs

This parameter defines the interval between synchronization epochs for NDB Cluster Replication. The default value is 100 milliseconds.

TimeBetweenEpochs is part of the implementation of “micro-GCPs”, which can be used to improve the performance of NDB Cluster Replication.

TimeBetweenEpochsTimeout

This parameter defines a timeout for synchronization epochs for NDB Cluster Replication. If a node fails to participate in a global checkpoint within the time determined by this parameter, the node is shut down. The default value is 0; in other words, the timeout is disabled.

TimeBetweenEpochsTimeout is part of the implementation of “micro-GCPs”, which can be used to improve the performance of NDB Cluster Replication.

The current value of this parameter and a warning are written to the cluster log whenever a GCP save takes longer than 1 minute or a GCP commit takes longer than 10 seconds.

Setting this parameter to zero has the effect of disabling GCP stops caused by save timeouts, commit timeouts, or both. The maximum possible value for this parameter is 256000 milliseconds.

MaxBufferedEpochs

The number of unprocessed epochs by which a subscribing node can lag behind. Exceeding this number causes a lagging subscriber to be disconnected.

The default value of 100 is sufficient for most normal operations. If a subscribing node does lag enough to cause disconnections, it is usually due to network or scheduling issues with regard to processes or threads. (In rare circumstances, the problem may be due to a bug in the NDB client.) It may be desirable to set the value lower than the default when epochs are longer.

Disconnection prevents client issues from affecting the data node service, running out of memory to buffer data, and eventually shutting down. Instead, only the client is affected as a result of the disconnect (by, for example gap events in the binary log), forcing the client to reconnect or restart the process.

MaxBufferedEpochBytes

The total number of bytes allocated for buffering epochs by this node.

TimeBetweenInactiveTransactionAbortCheck

Timeout handling is performed by checking a timer on each transaction once for every interval specified by this parameter. Thus, if this parameter is set to 1000 milliseconds, every transaction will be checked for timing out once per second.

The default value is 1000 milliseconds (1 second).

TransactionInactiveTimeout

This parameter states the maximum time that is permitted to lapse between operations in the same transaction before the transaction is aborted.

The default for this parameter is 4G (also the maximum). For a real-time database that needs to ensure that no transaction keeps locks for too long, this parameter should be set to a relatively small value. Setting it to 0 means that the application never times out. The unit is milliseconds.

TransactionDeadlockDetectionTimeout

When a node executes a query involving a transaction, the node waits for the other nodes in the cluster to respond before continuing. This parameter sets the amount of time that the transaction can spend executing within a data node, that is, the time that the transaction coordinator waits for each data node participating in the transaction to execute a request.

A failure to respond can occur for any of the following reasons:

This timeout parameter states how long the transaction coordinator waits for query execution by another node before aborting the transaction, and is important for both node failure handling and deadlock detection.

The default timeout value is 1200 milliseconds (1.2 seconds).

The minimum for this parameter is 50 milliseconds.

DiskSyncSize

This is the maximum number of bytes to store before flushing data to a local checkpoint file. This is done to prevent write buffering, which can impede performance significantly. This parameter is not intended to take the place of TimeBetweenLocalCheckpoints.

The default value is 4M (4 megabytes).

MaxDiskWriteSpeed

Set the maximum rate for writing to disk, in bytes per second, by local checkpoints and backup operations when no restarts (by this data node or any other data node) are taking place in this NDB Cluster.

For setting the maximum rate of disk writes allowed while this data node is restarting, use MaxDiskWriteSpeedOwnRestart. For setting the maximum rate of disk writes allowed while other data nodes are restarting, use MaxDiskWriteSpeedOtherNodeRestart. The minimum speed for disk writes by all LCPs and backup operations can be adjusted by setting MinDiskWriteSpeed.

MaxDiskWriteSpeedOtherNodeRestart

Set the maximum rate for writing to disk, in bytes per second, by local checkpoints and backup operations when one or more data nodes in this NDB Cluster are restarting, other than this node.

For setting the maximum rate of disk writes allowed while this data node is restarting, use MaxDiskWriteSpeedOwnRestart. For setting the maximum rate of disk writes allowed when no data nodes are restarting anywhere in the cluster, use MaxDiskWriteSpeed. The minimum speed for disk writes by all LCPs and backup operations can be adjusted by setting MinDiskWriteSpeed.

MaxDiskWriteSpeedOwnRestart

Set the maximum rate for writing to disk, in bytes per second, by local checkpoints and backup operations while this data node is restarting.

For setting the maximum rate of disk writes allowed while other data nodes are restarting, use MaxDiskWriteSpeedOtherNodeRestart. For setting the maximum rate of disk writes allowed when no data nodes are restarting anywhere in the cluster, use MaxDiskWriteSpeed. The minimum speed for disk writes by all LCPs and backup operations can be adjusted by setting MinDiskWriteSpeed.

MinDiskWriteSpeed

Set the minimum rate for writing to disk, in bytes per second, by local checkpoints and backup operations.

The maximum rates of disk writes allowed for LCPs and backups under various conditions are adjustable using the parameters MaxDiskWriteSpeed, MaxDiskWriteSpeedOwnRestart, and MaxDiskWriteSpeedOtherNodeRestart. See the descriptions of these parameters for more information.

ArbitrationTimeout

This parameter specifies how long data nodes wait for a response from the arbitrator to an arbitration message. If this is exceeded, the network is assumed to have split.

The default value is 7500 milliseconds (7.5 seconds).

Arbitration

The Arbitration parameter enables a choice of arbitration schemes, corresponding to one of 3 possible values for this parameter:

RestartSubscriberConnectTimeout

This parameter determines the time that a data node waits for subscribing API nodes to connect. Once this timeout expires, any “missing” API nodes are disconnected from the cluster. To disable this timeout, set RestartSubscriberConnectTimeout to 0.

While this parameter is specified in milliseconds, the timeout itself is resolved to the next-greatest whole second.

UndoIndexBuffer

The UNDO index buffer, whose size is set by this parameter, is used during local checkpoints. The NDB storage engine uses a recovery scheme based on checkpoint consistency in conjunction with an operational REDO log. To produce a consistent checkpoint without blocking the entire system for writes, UNDO logging is done while performing the local checkpoint. UNDO logging is activated on a single table fragment at a time. This optimization is possible because tables are stored entirely in main memory.

The UNDO index buffer is used for the updates on the primary key hash index. Inserts and deletes rearrange the hash index; the NDB storage engine writes UNDO log records that map all physical changes to an index page so that they can be undone at system restart. It also logs all active insert operations for each fragment at the start of a local checkpoint.

Reads and updates set lock bits and update a header in the hash index entry. These changes are handled by the page-writing algorithm to ensure that these operations need no UNDO logging.

This buffer is 2MB by default. The minimum value is 1MB, which is sufficient for most applications. For applications doing extremely large or numerous inserts and deletes together with large transactions and large primary keys, it may be necessary to increase the size of this buffer. If this buffer is too small, the NDB storage engine issues internal error code 677 (Index UNDO buffers overloaded).

UndoDataBuffer

This parameter sets the size of the UNDO data buffer, which performs a function similar to that of the UNDO index buffer, except the UNDO data buffer is used with regard to data memory rather than index memory. This buffer is used during the local checkpoint phase of a fragment for inserts, deletes, and updates.

Because UNDO log entries tend to grow larger as more operations are logged, this buffer is also larger than its index memory counterpart, with a default value of 16MB.

This amount of memory may be unnecessarily large for some applications. In such cases, it is possible to decrease this size to a minimum of 1MB.

It is rarely necessary to increase the size of this buffer. If there is such a need, it is a good idea to check whether the disks can actually handle the load caused by database update activity. A lack of sufficient disk space cannot be overcome by increasing the size of this buffer.

If this buffer is too small and gets congested, the NDB storage engine issues internal error code 891 (Data UNDO buffers overloaded).

RedoBuffer

All update activities also need to be logged. The REDO log makes it possible to replay these updates whenever the system is restarted. The NDB recovery algorithm uses a “fuzzy” checkpoint of the data together with the UNDO log, and then applies the REDO log to play back all changes up to the restoration point.

RedoBuffer sets the size of the buffer in which the REDO log is written. The default value is 32MB; the minimum value is 1MB.

If this buffer is too small, the NDB storage engine issues error code 1221 (REDO log buffers overloaded). For this reason, you should exercise care if you attempt to decrease the value of RedoBuffer as part of an online change in the cluster's configuration.

ndbmtd allocates a separate buffer for each LDM thread (see ThreadConfig). For example, with 4 LDM threads, an ndbmtd data node actually has 4 buffers and allocates RedoBuffer bytes to each one, for a total of 4 * RedoBuffer bytes.

EventLogBufferSize

Controls the size of the circular buffer used for NDB log events within data nodes.

LogLevelStartup

The reporting level for events generated during startup of the process.

The default level is 1.

LogLevelShutdown

The reporting level for events generated as part of graceful shutdown of a node.

The default level is 0.

LogLevelStatistic

The reporting level for statistical events such as number of primary key reads, number of updates, number of inserts, information relating to buffer usage, and so on.

The default level is 0.

LogLevelCheckpoint

The reporting level for events generated by local and global checkpoints.

The default level is 0.

LogLevelNodeRestart

The reporting level for events generated during node restart.

The default level is 0.

LogLevelConnection

The reporting level for events generated by connections between cluster nodes.

The default level is 0.

LogLevelError

The reporting level for events generated by errors and warnings by the cluster as a whole. These errors do not cause any node failure but are still considered worth reporting.

The default level is 0.

LogLevelCongestion

The reporting level for events generated by congestion. These errors do not cause node failure but are still considered worth reporting.

The default level is 0.

LogLevelInfo

The reporting level for events generated for information about the general state of the cluster.

The default level is 0.

MemReportFrequency

This parameter controls how often data node memory usage reports are recorded in the cluster log; it is an integer value representing the number of seconds between reports.

Each data node's data memory and index memory usage is logged as both a percentage and a number of 32 KB pages of DataMemory, as set in the config.ini file. For example, if DataMemory is equal to 100 MB, and a given data node is using 50 MB for data memory storage, the corresponding line in the cluster log might look like this:

2006-12-24 01:18:16 [MgmSrvr] INFO -- Node 2: Data usage is 50%(1280 32K pages of total 2560)

MemReportFrequency is not a required parameter. If used, it can be set for all cluster data nodes in the [ndbd default] section of config.ini, and can also be set or overridden for individual data nodes in the corresponding [ndbd] sections of the configuration file. The minimum value—which is also the default value—is 0, in which case memory reports are logged only when memory usage reaches certain percentages (80%, 90%, and 100%), as mentioned in the discussion of statistics events in Section 22.5.6.2, “NDB Cluster Log Events”.

StartupStatusReportFrequency

When a data node is started with the --initial, it initializes the redo log file during Start Phase 4 (see Section 22.5.1, “Summary of NDB Cluster Start Phases”). When very large values are set for NoOfFragmentLogFiles, FragmentLogFileSize, or both, this initialization can take a long time.You can force reports on the progress of this process to be logged periodically, by means of the StartupStatusReportFrequency configuration parameter. In this case, progress is reported in the cluster log, in terms of both the number of files and the amount of space that have been initialized, as shown here:

2009-06-20 16:39:23 [MgmSrvr] INFO -- Node 1: Local redo log file initialization status:
#Total files: 80, Completed: 60
#Total MBytes: 20480, Completed: 15557
2009-06-20 16:39:23 [MgmSrvr] INFO -- Node 2: Local redo log file initialization status:
#Total files: 80, Completed: 60
#Total MBytes: 20480, Completed: 15570

These reports are logged each StartupStatusReportFrequency seconds during Start Phase 4. If StartupStatusReportFrequency is 0 (the default), then reports are written to the cluster log only when at the beginning and at the completion of the redo log file initialization process.

DictTrace

It is possible to cause logging of traces for events generated by creating and dropping tables using DictTrace. This parameter is useful only in debugging NDB kernel code. DictTrace takes an integer value. 0 is the default, and means no logging is performed; 1 enables trace logging, and 2 enables logging of additional DBDICT debugging output.

WatchdogImmediateKill

You can cause threads to be killed immediately whenever watchdog issues occur by enabling the WatchdogImmediateKill data node configuration parameter. This parameter should be used only when debugging or troubleshooting, to obtain trace files reporting exactly what was occurring the instant that execution ceased.

BackupDataBufferSize

In creating a backup, there are two buffers used for sending data to the disk. The backup data buffer is used to fill in data recorded by scanning a node's tables. Once this buffer has been filled to the level specified as BackupWriteSize, the pages are sent to disk. While flushing data to disk, the backup process can continue filling this buffer until it runs out of space. When this happens, the backup process pauses the scan and waits until some disk writes have completed freeing up memory so that scanning may continue.

The default value for this parameter is 16MB. The minimum is 512K.

BackupDiskWriteSpeedPct

During normal operation, data nodes attempt to maximize the disk write speed used for local checkpoints and backups while remaining within the bounds set by MinDiskWriteSpeed and MaxDiskWriteSpeed. Disk write throttling gives each LDM thread an equal share of the total budget. This allows parallel LCPs to take place without exceeding the disk I/O budget. Because a backup is executed by only one LDM thread, this effectively caused a budget cut, resulting in longer backup completion times, and—if the rate of change is sufficiently high—in failure to complete the backup when the backup log buffer fill rate is higher than the achievable write rate.

This problem can be addressed by using the BackupDiskWriteSpeedPct configuration parameter, which takes a value in the range 0-90 (inclusive) which is interpreted as the percentage of the node's maximum write rate budget that is reserved prior to sharing out the remainder of the budget among LDM threads for LCPs. The LDM thread running the backup receives the whole write rate budget for the backup, plus its (reduced) share of the write rate budget for local checkpoints.

The default value for this parameter is 50 (interpreted as 50%).

BackupLogBufferSize

The backup log buffer fulfills a role similar to that played by the backup data buffer, except that it is used for generating a log of all table writes made during execution of the backup. The same principles apply for writing these pages as with the backup data buffer, except that when there is no more space in the backup log buffer, the backup fails. For that reason, the size of the backup log buffer must be large enough to handle the load caused by write activities while the backup is being made. See Section 22.5.3.3, “Configuration for NDB Cluster Backups”.

The default value for this parameter should be sufficient for most applications. In fact, it is more likely for a backup failure to be caused by insufficient disk write speed than it is for the backup log buffer to become full. If the disk subsystem is not configured for the write load caused by applications, the cluster is unlikely to be able to perform the desired operations.

It is preferable to configure cluster nodes in such a manner that the processor becomes the bottleneck rather than the disks or the network connections.

The default value for this parameter is 16MB.

BackupMemory

This parameter is deprecated, and subject to removal in a future version of NDB Cluster. Any setting made for it is ignored.

BackupReportFrequency

This parameter controls how often backup status reports are issued in the management client during a backup, as well as how often such reports are written to the cluster log (provided cluster event logging is configured to permit it—see Logging and checkpointing). BackupReportFrequency represents the time in seconds between backup status reports.

The default value is 0.

BackupWriteSize

This parameter specifies the default size of messages written to disk by the backup log and backup data buffers.

The default value for this parameter is 256KB.

BackupMaxWriteSize

This parameter specifies the maximum size of messages written to disk by the backup log and backup data buffers.

The default value for this parameter is 1MB.

BackupDataBufferSize >= BackupWriteSize + 188KB

BackupLogBufferSize >= BackupWriteSize + 16KB

BackupMaxWriteSize >= BackupWriteSize

LockExecuteThreadToCPU

When used with ndbd, this parameter (now a string) specifies the ID of the CPU assigned to handle the NDBCLUSTER execution thread. When used with ndbmtd, the value of this parameter is a comma-separated list of CPU IDs assigned to handle execution threads. Each CPU ID in the list should be an integer in the range 0 to 65535 (inclusive).

The number of IDs specified should match the number of execution threads determined by MaxNoOfExecutionThreads. However, there is no guarantee that threads are assigned to CPUs in any given order when using this parameter. You can obtain more finely-grained control of this type using ThreadConfig.

LockExecuteThreadToCPU has no default value.

LockMaintThreadsToCPU

This parameter specifies the ID of the CPU assigned to handle NDBCLUSTER maintenance threads.

The value of this parameter is an integer in the range 0 to 65535 (inclusive). There is no default value.

RealtimeScheduler

Setting this parameter to 1 enables real-time scheduling of data node threads.

The default is 0 (scheduling disabled).

SchedulerExecutionTimer

This parameter specifies the time in microseconds for threads to be executed in the scheduler before being sent. Setting it to 0 minimizes the response time; to achieve higher throughput, you can increase the value at the expense of longer response times.

The default is 50 μsec, which our testing shows to increase throughput slightly in high-load cases without materially delaying requests.

SchedulerResponsiveness

Set the balance in the NDB scheduler between speed and throughput. This parameter takes an integer whose value is in the range 0-10 inclusive, with 5 as the default. Higher values provide better response times relative to throughput. Lower values provide increased throughput at the expense of longer response times.

SchedulerSpinTimer

This parameter specifies the time in microseconds for threads to be executed in the scheduler before sleeping.

The default value is 0.

BuildIndexThreads

This parameter determines the number of threads to create when rebuilding ordered indexes during a system or node start, as well as when running ndb_restore --rebuild-indexes. It is supported only when there is more than one fragment for the table per data node (for example, when COMMENT="NDB_TABLE=PARTITION_BALANCE=FOR_RA_BY_LDM_X_2" is used with CREATE TABLE).

Setting this parameter to 0 (the default) disables multithreaded building of ordered indexes.

This parameter is supported when using ndbd or ndbmtd.

You can enable multithreaded builds during data node initial restarts by setting the TwoPassInitialNodeRestartCopy data node configuration parameter to TRUE.

TwoPassInitialNodeRestartCopy

Multithreaded building of ordered indexes can be enabled for initial restarts of data nodes by setting this configuration parameter to true (the default value), which enables two-pass copying of data during initial node restarts.

You must also set BuildIndexThreads to a nonzero value.

Numa

This parameter determines whether Non-Uniform Memory Access (NUMA) is controlled by the operating system or by the data node process, whether the data node uses ndbd or ndbmtd. By default, NDB attempts to use an interleaved NUMA memory allocation policy on any data node where the host operating system provides NUMA support.

Setting Numa = 0 means that the datanode process does not itself attempt to set a policy for memory allocation, and permits this behavior to be determined by the operating system, which may be further guided by the separate numactl tool. That is, Numa = 0 yields the system default behavior, which can be customised by numactl. For many Linux systems, the system default behavior is to allocate socket-local memory to any given process at allocation time. This can be problematic when using ndbmtd; this is because nbdmtd allocates all memory at startup, leading to an imbalance, giving different access speeds for different sockets, especially when locking pages in main memory.

Setting Numa = 1 means that the data node process uses libnuma to request interleaved memory allocation. (This can also be accomplished manually, on the operating system level, using numactl.) Using interleaved allocation in effect tells the data node process to ignore non-uniform memory access but does not attempt to take any advantage of fast local memory; instead, the data node process tries to avoid imbalances due to slow remote memory. If interleaved allocation is not desired, set Numa to 0 so that the desired behavior can be determined on the operating system level.

The Numa configuration parameter is supported only on Linux systems where libnuma.so is available.

MaxNoOfExecutionThreads

This parameter directly controls the number of execution threads used by ndbmtd, up to a maximum of 72. Although this parameter is set in [ndbd] or [ndbd default] sections of the config.ini file, it is exclusive to ndbmtd and does not apply to ndbd.

Setting MaxNoOfExecutionThreads sets the number of threads for each type as determined by a matrix in the file storage/ndb/src/kernel/vm/mt_thr_config.cpp. This table shows these numbers of threads for possible values of MaxNoOfExecutionThreads.