Rate limiter manages the rate at which requests are scheduled on model instances by Triton. The rate limiter operates across all models loaded in Triton to allow cross-model prioritization.
In absence of rate limiting (--rate-limit=off), Triton schedules execution of a request (or set of requests when using dynamic batching) as soon as a model instance is available. This behavior is typically best suited for performance. However, there can be cases where running all the models simultaneously places excessive load on the server. For instance, model execution on some frameworks dynamically allocate memory. Running all such models simultaneously may lead to system going out-of-memory.
Rate limiter allows to postpone the inference execution on some model instances such that not all of them runs simultaneously. The model priorities are used to decide which model instance to schedule next.
To enable rate limiting users must set --rate-limit option when
launching tritonserver. For more information, consult usage of
the option emitted by tritonserver --help.
The rate limiter is controlled by the rate limiter configuration given for each model instance, as described in rate limiter configuration. The rate limiter configuration includes resources and priority for the model instances defined by the instance group.
Resources are identified by a unique name and a count indicating the number of copies of the resource. By default, model instance uses no rate-limiter resources. By listing a resource/count the model instance indicates that it requires that many resources to be available on the model instance device before it can be allowed to execute. When under execution the specified many resources are allocated to the model instance only to be released when the execution is over. The available number of resource copies are, by default, the max across all model instances that list that resource. For example, assume three loaded model instances A, B and C each specifying the following resource requirements for a single device:
A: [R1: 4, R2: 4]
B: [R2: 5, R3: 10, R4: 5]
C: [R1: 1, R3: 7, R4: 2]
By default, based on those model instance requirements, the server will create the following resources with the indicated copies:
R1: 4
R2: 5
R3: 10
R4: 5
These values ensure that all model instances can be successfully
scheduled. The default for a resource can be overridden by giving
it explicitly on command-line using --rate-limit-resource option.
tritonserver --help will provide with more detailed usage
instructions.
By default, the available resource copies are per-device and resource
requirements for a model instance are enforced against corresponding
resources associated with the device where the model instance runs.
The --rate-limit-resource allows users to provide different resource
copies to different devices. Rate limiter can also handle global
resources. Instead of creating resource copies per-device, a global
resource will have a single copy all across the system.
Rate limiter depends upon the model configuration to determine whether the resource is global or not. See resources for more details on how to specify them in model configuration.
For tritonserver, running on a two device machine, invoked with
--rate-limit-resource=R1:10 --rate-limit-resource=R2:5:0 --rate-limit-resource=R2:8:1 --rate-limit-resource=R3:2
, available resource copies are:
GLOBAL => [R3: 2]
DEVICE 0 => [R1: 10, R2: 5]
DEVICE 1 => [R1: 10, R2: 8]
where R3 appears as a global resource in one of the loaded model.
In a resource constrained system, there will be a contention for the resources among model instances to execute their inference requests. Priority setting helps determining which model instance to select for next execution. See priority for more information.