Formerly, the NDB kernel employed a send buffer whose size was fixed at 2MB for each node in the cluster, this buffer being allocated when the node started. Because the size of this buffer could not be changed after the cluster was started, it was necessary to make it large enough in advance to accommodate the maximum possible load on any transporter socket. However, this was an inefficient use of memory, since much of it often went unused, and could result in large amounts of resources being wasted when scaling up to many API nodes.
This problem was eventually solved (in NDB Cluster 7.0) by employing a unified send buffer whose memory is allocated dynamically from a pool shared by all transporters. This means that the size of the send buffer can be adjusted as necessary. Configuration of the unified send buffer can accomplished by setting the following parameters:
TotalSendBufferMemory. This parameter can be set for all types of NDB Cluster nodes—that is, it can be set in the
[mysql]) sections of the
config.inifile. It represents the total amount of memory (in bytes) to be allocated by each node for which it is set for use among all configured transporters. If set, its minimum is 256KB; the maximum is 4294967039.
To be backward-compatible with existing configurations, this parameter takes as its default value the sum of the maximum send buffer sizes of all configured transporters, plus an additional 32KB (one page) per transporter. The maximum depends on the type of transporter, as shown in the following table:
Transporter Maximum Send Buffer Size (bytes) TCP
SendBufferMemory(default = 2M)
SendLimit(default = 8K) plus 16K
This enables existing configurations to function in close to the same way as they did with NDB Cluster 6.3 and earlier, with the same amount of memory and send buffer space available to each transporter. However, memory that is unused by one transporter is not available to other transporters.
OverloadLimit. This parameter is used in the
[tcp]section, and denotes the amount of unsent data (in bytes) that must be present in the send buffer before the connection is considered overloaded. When such an overload condition occurs, transactions that affect the overloaded connection fail with NDB API Error 1218 (Send Buffers overloaded in NDB kernel) until the overload status passes. The default value is 0, in which case the effective overload limit is calculated as
SendBufferMemory * 0.8for a given connection. The maximum value for this parameter is 4G.
SendBufferMemory. This value denotes a hard limit for the amount of memory that may be used by a single transporter out of the entire pool specified by
TotalSendBufferMemory. However, the sum of
SendBufferMemoryfor all configured transporters may be greater than the
TotalSendBufferMemorythat is set for a given node. This is a way to save memory when many nodes are in use, as long as the maximum amount of memory is never required by all transporters at the same time.
ReservedSendBufferMemory. This optional data node parameter, if set, gives an amount of memory (in bytes) that is reserved for connections between data nodes; this memory is not allocated to send buffers used for communications with management servers or API nodes. This provides a way to protect the cluster against misbehaving API nodes that use excess send memory and thus cause failures in communications internally in the NDB kernel. If set, its the minimum permitted value for this parameters is 256KB; the maximum is 4294967039.
You can use the
ndbinfo.transporters table to
monitor send buffer memory usage, and to detect slowdown and
overload conditions that can adversely affect performance.