NDB management client's
CLUSTERLOG STATISTICS command can provide a
number of useful statistics in its output. Counters providing
information about the state of the cluster are updated at
5-second reporting intervals by the transaction coordinator (TC)
and the local query handler (LQH), and written to the cluster
Transaction coordinator statistics. Each transaction has one transaction coordinator, which is chosen by one of the following methods:
In a round-robin fashion
By communication proximity
You can determine which TC selection method is used for
transactions started from a given SQL node using the
system variable. For more information, see
Section 188.8.131.52, “MySQL Cluster System Variables”.
All operations within the same transaction use the same transaction coordinator, which reports the following statistics:
Trans count. This is the number transactions started in the last interval using this TC as the transaction coordinator. Any of these transactions may have committed, have been aborted, or remain uncommitted at the end of the reporting interval.
Transactions do not migrate between TCs.
This is the number of transactions using this TC as the
transaction coordinator that were committed in the last
reporting interval. Because some transactions committed in
this reporting interval may have started in a previous
reporting interval, it is possible for
count to be greater than
Read count. This is the number of primary key read operations using this TC as the transaction coordinator that were started in the last reporting interval, including simple reads. This count also includes reads performed as part of unique index operations. A unique index read operation generates 2 primary key read operations—1 for the hidden unique index table, and 1 for the table on which the read takes place.
Simple read count.
This is the number of simple read operations using this TC
as the transaction coordinator that were started in the
last reporting interval. This is a subset of
count. Because the value of
count is incremented at a different point in
Read count, it can lag behind
Read count slightly, so it is
Simple read count is
not equal to
Read count for a given
reporting interval, even if all reads made during that
time were in fact simple reads.
Write count. This is the number of primary key write operations using this TC as the transaction coordinator that were started in the last reporting interval. This includes all inserts, updates, writes and deletes, as well as writes performed as part of unique index operations.
A unique index update operation can generate multiple PK read and write operations on the index table and on the base table.
This is the number of 32-bit data words received in the
last reporting interval for primary key operations using
this TC as the transaction coordinator. For reads, this is
proportional to the number of columns requested. For
inserts and updates, this is proportional to the number of
columns written, and the size of their data. For delete
operations, this is usually zero. Unique index operations
generate multiple PK operations and so increase this
count. However, data words sent to describe the PK
operation itself, and the key information sent, are
not counted here. Attribute
information sent to describe columns to read for scans, or
to describe ScanFilters, is also not counted in
This is the number of primary key or scan operations using
this TC as the transaction coordinator that were started
during the last reporting interval but that were not
completed. Operations increment this counter when they are
started and decrement it when they are completed; this
occurs after the transaction commits. Dirty reads and
writes—as well as failed operations—decrement
this counter. The maximum value that
Operations can have is the maximum number of
operations that a TC block can support; currently, this is
(2 * MaxNoOfConcurrentOperations) + 16 +
MaxNoOfConcurrentTransactions. (For more
information about these configuration parameters, see the
Transaction Parameters section of
Section 184.108.40.206, “Defining MySQL Cluster Data Nodes”.)
This is the number of transactions using this TC as the
transaction coordinator that were aborted during the last
reporting interval. Because some transactions that were
aborted in the last reporting interval may have started in
a previous reporting interval,
count can sometimes be greater than
Scans. This is the number of table scans using this TC as the transaction coordinator that were started during the last reporting interval. This does not include range scans (that is, ordered index scans).
Range scans. This is the number of ordered index scans using this TC as the transaction coordinator that were started in the last reporting interval.
Local query handler statistics (Operations). There is 1 cluster event per local query handler block (that is, 1 per data node process). Operations are recorded in the LQH where the data they are operating on resides.
A single transaction may operate on data stored in multiple LQH blocks.
Operations statistic provides the number
of local operations performed by this LQH block in the last
reporting interval, and includes all types of read and write
operations (insert, update, write, and delete operations). This
also includes operations used to replicate writes—for
example, in a 2-replica cluster, the write to the primary
replica is recorded in the primary LQH, and the write to the
backup will be recorded in the backup LQH. Unique key operations
may result in multiple local operations; however, this does
not include local operations generated as a
result of a table scan or ordered index scan, which are not
Process scheduler statistics. In addition to the statistics reported by the transaction coordinator and local query handler, each ndbd process has a scheduler which also provides useful metrics relating to the performance of a MySQL Cluster. This scheduler runs in an infinite loop; during each loop the scheduler performs the following tasks:
Read any incoming messages from sockets into a job buffer.
Check whether there are any timed messages to be executed; if so, put these into the job buffer as well.
Execute (in a loop) any messages in the job buffer.
Send any distributed messages that were generated by executing the messages in the job buffer.
Wait for any new incoming messages.
Process scheduler statistics include the following:
Mean Loop Counter. This is the number of loops executed in the third step from the preceding list. This statistic increases in size as the utilization of the TCP/IP buffer improves. You can use this to monitor changes in performance as you add new data node processes.
Mean send size and Mean receive size. These statistics enable you to gauge the efficiency of, respectively writes and reads between nodes. The values are given in bytes. Higher values mean a lower cost per byte sent or received; the maximum value is 64K.
To cause all cluster log statistics to be logged, you can use
the following command in the
ALL CLUSTERLOG STATISTICS=15
Setting the threshold for
STATISTICS to 15
causes the cluster log to become very verbose, and to grow
quite rapidly in size, in direct proportion to the number of
cluster nodes and the amount of activity in the MySQL Cluster.
For more information about MySQL Cluster management client commands relating to logging and reporting, see Section 220.127.116.11, “MySQL Cluster Logging Management Commands”.