25.4.3.12 NDB Cluster Shared-Memory Connections (original) (raw)

25.4.3.12 NDB Cluster Shared-Memory Connections

Communications between NDB cluster nodes are normally handled using TCP/IP. The shared memory (SHM) transporter is distinguished by the fact that signals are transmitted by writing in memory rather than on a socket. The shared-memory transporter (SHM) can improve performance by negating up to 20% of the overhead required by a TCP connection when running an API node (usually an SQL node) and a data node together on the same host. You can enable a shared memory connection in either of the two ways listed here:

Suppose a cluster is running a data node which has node ID 1 and an SQL node having node ID 51 on the same host computer at 10.0.0.1. To enable an SHM connection between these two nodes, all that is necessary is to insure that the following entries are included in the cluster configuration file:

[ndbd]
NodeId=1
HostName=10.0.0.1
UseShm=1

[mysqld]
NodeId=51
HostName=10.0.0.1

Important

The two entries just shown are in addition to any other entries and parameter settings needed by the cluster. A more complete example is shown later in this section.

Before starting data nodes that use SHM connections, it is also necessary to make sure that the operating system on each computer hosting such a data node has sufficient memory allocated to shared memory segments. See the documentation for your operating platform for information regarding this. In setups where multiple hosts are each running a data node and an API node, it is possible to enable shared memory on all such hosts by setting UseShm in the [ndbd default] section of the configuration file. This is shown in the example later in this section.

While not strictly required, tuning for all SHM connections in the cluster can be done by setting one or more of the following parameters in the [shm default] section of the cluster configuration (config.ini) file:

This example shows a simple setup with SHM connections defined on multiple hosts, in an NDB Cluster using 3 computers listed here by host name, hosting the node types shown:

  1. 10.0.0.0: The management server
  2. 10.0.0.1: A data node and an SQL node
  3. 10.0.0.2: A data node and an SQL node

In this scenario, each data node communicates with both the management server and the other data node using TCP transporters; each SQL node uses a shared memory transporter to communicate with the data nodes that is local to it, and a TCP transporter to communicate with the remote data node. A basic configuration reflecting this setup is enabled by the config.ini file whose contents are shown here:

[ndbd default]
DataDir=/path/to/datadir
UseShm=1

[shm default]
ShmSize=8M
ShmSpintime=200
SendBufferMemory=4M

[tcp default]
SendBufferMemory=8M

[ndb_mgmd]
NodeId=49
Hostname=10.0.0.0
DataDir=/path/to/datadir

[ndbd]
NodeId=1
Hostname=10.0.0.1
DataDir=/path/to/datadir

[ndbd]
NodeId=2
Hostname=10.0.0.2
DataDir=/path/to/datadir

[mysqld]
NodeId=51
Hostname=10.0.0.1

[mysqld]
NodeId=52
Hostname=10.0.0.2

[api]
[api]

Parameters affecting all shared memory transporters are set in the [shm default] section; these can be overridden on a per-connection basis in one or more[shm] sections. Each such section must be associated with a given SHM connection usingNodeId1 andNodeId2; the values required for these parameters are the node IDs of the two nodes connected by the transporter. You can also identify the nodes by host name using HostName1 and HostName2, but these parameters are not required.

The API nodes for which no host names are set use the TCP transporter to communicate with data nodes independent of the hosts on which they are started; the parameters and values set in the [tcp default] section of the configuration file apply to all TCP transporters in the cluster.

For optimum performance, you can define a spin time for the SHM transporter (ShmSpinTime parameter); this affects both the data node receiver thread and the poll owner (receive thread or user thread) inNDB.

Restart types. Information about the restart types used by the parameter descriptions in this section is shown in the following table:

Table 25.22 NDB Cluster restart types

Symbol Restart Type Description
N Node The parameter can be updated using a rolling restart (seeSection 25.6.5, “Performing a Rolling Restart of an NDB Cluster”)
S System All cluster nodes must be shut down completely, then restarted, to effect a change in this parameter
I Initial Data nodes must be restarted using the--initial option