Node scheduler properties#
The target value for the total number of splits that can be running for each worker node, assuming all splits have the standard split weight.
Using a higher value is recommended, if queries are submitted in large batches (e.g., running a large group of reports periodically), or for connectors that produce many splits that complete quickly but do not support assigning split weight values to express that to the split scheduler. Increasing this value may improve query latency, by ensuring that the workers have enough splits to keep them fully utilized.
When connectors do support weight based split scheduling, the number of splits assigned will depend on the weight of the individual splits. If splits are small, more of them are allowed to be assigned to each worker to compensate.
Setting this too high wastes memory and may result in lower performance due to splits not being balanced across workers. Ideally, it should be set such that there is always at least one split waiting to be processed, but not higher.
The minimum number of outstanding splits with the standard split weight guaranteed to be scheduled on a node (even when the node is already at the limit for total number of splits) for a single task given the task has remaining splits to process. Allowing a minimum number of splits per stage is required to prevent starvation and deadlocks.
This value must be smaller or equal than
node-scheduler.max-splits-per-node, is usually increased for the same reasons,
and has similar drawbacks if set too high.
The maximum number of outstanding splits with the standard split weight guaranteed to be scheduled on a node (even when the node
is already at the limit for total number of splits) for a single task given the task has remaining splits to process.
Split queue size is adjusted dynamically during split scheduling and cannot exceed
Split queue size per task will be adjusted upward if node processes splits faster than it receives them.
Usually increased for the same reasons as
node-scheduler.max-splits-per-node, with smaller drawbacks
if set too high.
Only applies for
uniform scheduler policy.
The minimum number of candidate nodes that are evaluated by the node scheduler when choosing the target node for a split. Setting this value too low may prevent splits from being properly balanced across all worker nodes. Setting it too high may increase query latency and increase CPU usage on the coordinator.
Sets the node scheduler policy to use when scheduling splits.
to schedule splits on the host where the data is located, while maintaining a uniform
distribution across all hosts.
topology tries to schedule splits according to
the topology distance between nodes and splits. It is recommended to use
for clusters where distributed storage runs on the same nodes as Trino workers.
A comma-separated string describing the meaning of each segment of a network location.
For example, setting
region,rack,machine means a network location contains three segments.
Sets the network topology type. To use this option,
must be set to
flat: the topology has only one segment, with one value for each machine.
file: the topology is loaded from a file using the properties
node-scheduler.network-topology.refresh-perioddescribed in the following sections.
subnet: the topology is derived based on subnet configuration provided through properties
node-scheduler.network-topology.subnet.ip-address-protocoldescribed in the following sections.
File based network topology#
Load the network topology from a file. To use this option,
must be set to
file. Each line contains a mapping between a host name and a
network location, separated by whitespace. Network location must begin with a leading
/ and segments are separated by a
192.168.0.1 /region1/rack1/machine1 192.168.0.2 /region1/rack1/machine2 hdfs01.example.com /region2/rack2/machine3
Controls how often the network topology file is reloaded. To use this option,
node-scheduler.network-topology.type must be set to
Subnet based network topology#
Sets the IP address protocol to be used for computing subnet based
topology. To use this option,
be set to
A comma-separated list of integer values defining CIDR prefix
lengths for subnet masks. The prefix lengths must be in increasing order. The
maximum prefix length values for IPv4 and IPv6 protocols are 32 and 128
respectively. To use this option,
be set to
For example, the value
24,25,27 for this property with IPv4 protocol means
that masks applied on the IP address to compute location segments are
255.255.255.224. So the segments
created for an address