Metrics

Triton provides Prometheus metrics indicating GPU and request statistics. By default, these metrics are available at http://localhost:8002/metrics. The metrics are only available by accessing the endpoint, and are not pushed or published to any remote server. The metric format is plain text so you can view them directly, for example:

$ curl localhost:8002/metrics

The tritonserver --allow-metrics=false option can be used to disable all metric reporting, while the --allow-gpu-metrics=false and --allow-cpu-metrics=false can be used to disable just the GPU and CPU metrics respectively.

The --metrics-port option can be used to select a different port. By default, Triton reuses the --http-address option for the metrics endpoint and binds the http and metrics endpoints to the same specific address when http service is enabled. If http service is not enabled, the metric address will bind to 0.0.0.0 by default. To uniquely specify the metric endpoint, --metrics-address option can be used. See the tritonserver --help output for more info on these CLI options.

To change the interval at which metrics are polled/updated, see the --metrics-interval-ms flag. Metrics that are updated "Per Request" are unaffected by this interval setting. This interval only applies to metrics that are designated as "Per Interval" in the tables of each section below:

• Inference Request Metrics
• GPU Metrics
• CPU Metrics
• Pinned Memory Metrics
• Response Cache Metrics
• Custom Metrics

Inference Request Metrics

Counts

For models that do not support batching, Request Count, Inference Count and Execution Count will be equal, indicating that each inference request is executed separately.

For models that support batching, the count metrics can be interpreted to determine average batch size as Inference Count / Execution Count. The count metrics are illustrated by the following examples:

• Client sends a single batch-1 inference request. Request Count = 1, Inference Count = 1, Execution Count = 1.

• Client sends a single batch-8 inference request. Request Count = 1, Inference Count = 8, Execution Count = 1.

• Client sends 2 requests: batch-1 and batch-8. Dynamic batcher is not enabled for the model. Request Count = 2, Inference Count = 9, Execution Count = 2.

• Client sends 2 requests: batch-1 and batch-1. Dynamic batcher is enabled for the model and the 2 requests are dynamically batched by the server. Request Count = 2, Inference Count = 2, Execution Count = 1.

• Client sends 2 requests: batch-1 and batch-8. Dynamic batcher is enabled for the model and the 2 requests are dynamically batched by the server. Request Count = 2, Inference Count = 9, Execution Count = 1.

Failure Count Categories

Note

Ensemble failure metrics will reflect the failure counts of their composing models as well as the parent model, but currently do not capture the same granularity for the "reason" label and will default to the "OTHER" reason.

For example, if EnsembleA contains ModelA, and ModelA experiences a failed request due to a queue/backlog timeout in the scheduler, ModelA will have a failed request metric reflecting reason=REJECTED and count=1. Additionally, EnsembleA will have a failed request metric reflecting reason=OTHER and count=2. The count=2 reflects 1 from the internally failed request captured by ModelA, as well as 1 from the failed top-level request sent to EnsembleA by the user/client. The reason=OTHER reflects that fact that the ensemble doesn't currently capture the specific reason why ModelA's request failed at this time.

Pending Request Count (Queue Size) Per-Model

The Pending Request Count reflects the number of requests that have been received by Triton core via TRITONSERVER_InferAsync, but have not yet started execution by a backend model instance (TRITONBACKEND_ModelInstanceExecute).

For all intents and purposes, the "pending request count" and "queue size" per-model can be used interchangeably, and the number reflected in the metric should intuitively represent the number of requests that are not currently being executed by any model instances. In simple terms, if you send a 100 requests to a model that can only handle 5 requests concurrently, then you should see a pending count of 95 for that model in most cases.

For those interested in more technical details, the term "pending request count" is a bit more accurate than "queue size" because Triton is highly configurable, and there are many places in Triton that a request be considered pending rather than a single queue. Some of the most common will be called out below:

• Default Scheduler backlogs any requests not currently executing.
  • Assuming 1 available model instance with the default scheduler settings, and 10 requests are sent in rapid succession.
  • The 1st request should be picked up for execution immediately, and the remaining 9 requests should be considered pending for this model, until the 1st request is finished. Afterwards, the next request should be picked up and the pending count should be decremented to 8, and so on until all requests are finished and the pending count is 0.
• Dynamic Batcher queue for dynamically creating batches from requests.
  • Assuming 1 available model instance with the dynamic batch scheduler configured with max_batch_size: 4 and a sufficiently large max_queue_delay_microseconds (or queue of requests), and 10 requests are sent in rapid succession.
  • The first 4 requests, or as large of a batch the scheduler could form, should be picked up for execution immediately, and the remaining 6 requests should be considered pending. After the batch finishes, the next batch should be picked up, decrementing the pending count again to 2 pending. Then finally since only 2 requests remain, the final 2 requests will be batched and picked up by the backend, decrementing the pending count to 0.
• Sequence Batcher queues and backlogs for ongoing sequence requests, some may be assigned sequence slots, some may not.
  • Sequence Batchers of both strategies (direct and oldest) will have pending counts that generally follow the same trend as the dynamic batching description above. The sequence batchers will immediately execute as many requests in a batch as it can based on the model/scheduler config settings, and any further requests will be considered pending until the previous batch finishes and the next batch can start.
• Rate Limiter queues for prepared batches of requests.
  • When rate limiting is enabled, requests can be held back from execution to satisfy the rate limit constraints that were configured.

There are some places where a request would not be considered pending:

• Ensemble Scheduler
  • The Ensemble Scheduler almost immediately enqueues any requests it receives into the composing model schedulers at the first step in the ensemble. Therefore, the requests could be considered pending by the composing model scheduler's, however from the ensemble's perspective, these requests have been scheduled.
• Frontends (HTTP/GRPC Servers)
  • Any requests sent from a client to a frontend server in-front of Triton may spend some time in the corresponding server's code mapping protocol-specific metadata to Triton metadata. Though this time is generally brief, it will not be considered pending from Triton's perspective until Triton core has received the request from the frontend.

Latencies

Starting in 23.04, Triton exposes the ability to choose the types of metrics that are published through the --metrics-config CLI options.

Counters

By default, the following Counter metrics are used for latencies:

To disable these metrics specifically, you can set --metrics-config counter_latencies=false

Histograms

Note

The following Histogram feature is experimental for the time being and may be subject to change based on user feedback.

By default, the following Histogram metrics are used for latencies:

To enable these metrics specifically, you can set --metrics-config histogram_latencies=true

Each histogram above is composed of several sub-metrics. For each histogram metric, there is a set of le (less than or equal to) thresholds tracking the counter for each bucket. Additionally, there are _count and _sum metrics that aggregate the count and observed values for each. For example, see the following information exposed by the "Time to First Response" histogram metrics:

# HELP nv_first_response_histogram_ms Duration from request to first response in milliseconds
# TYPE nv_first_response_histogram_ms histogram
nv_inference_first_response_histogram_ms_count{model="my_model",version="1"} 37
nv_inference_first_response_histogram_ms_sum{model="my_model",version="1"} 10771
nv_inference_first_response_histogram_ms{model="my_model",version="1", le="100"} 8
nv_inference_first_response_histogram_ms{model="my_model",version="1", le="500"} 30
nv_inference_first_response_histogram_ms{model="my_model",version="1", le="2000"} 36
nv_inference_first_response_histogram_ms{model="my_model",version="1", le="5000"} 37
nv_inference_first_response_histogram_ms{model="my_model",version="1", le="+Inf"} 37

Triton initializes histograms with default buckets for each, as shown above. Buckets can be overridden per family by specifying model_metrics in the model configuration. For example:

// config.pbtxt
model_metrics {
  metric_control: [
    {
      metric_identifier: {
        family: "nv_inference_first_response_histogram_ms"
      }
      histogram_options: {
        buckets: [ 1, 2, 4, 8 ]
      }
    }
  ]
}

Note

To apply changes to metric options dynamically, the model must be completely unloaded and then reloaded for the updates to take effect.

Currently, the following histogram families support custom buckets.

nv_inference_first_response_histogram_ms  // Time to First Response

Summaries

Note

The following Summary feature is experimental for the time being and may be subject to change based on user feedback.

To get configurable quantiles over a sliding time window, Triton supports a set a Summary metrics for latencies as well. These metrics are disabled by default, but can be enabled by setting --metrics-config summary_latencies=true.

For more information on how the quantiles are calculated, see this explanation.

The following summary metrics are available:

Each summary above is actually composed of several sub-metrics. For each metric, there is a set of quantile metrics tracking the latency for each quantile. Additionally, there are _count and _sum metrics that aggregate the count and observed values for each. For example, see the following information exposed by the Inference Queue Summary metrics:

# HELP nv_inference_queue_summary_us Summary of inference queuing duration in microseconds (includes cached requests)
# TYPE nv_inference_queue_summary_us summary
nv_inference_queue_summary_us_count{model="my_model",version="1"} 161
nv_inference_queue_summary_us_sum{model="my_model",version="1"} 11110
nv_inference_queue_summary_us{model="my_model",version="1",quantile="0.5"} 55
nv_inference_queue_summary_us{model="my_model",version="1",quantile="0.9"} 97
nv_inference_queue_summary_us{model="my_model",version="1",quantile="0.95"} 98
nv_inference_queue_summary_us{model="my_model",version="1",quantile="0.99"} 101
nv_inference_queue_summary_us{model="my_model",version="1",quantile="0.999"} 101

The count and sum for the summary above show that stats have been recorded for 161 requests, and took a combined total of 11110 microseconds. The _count and _sum of a summary should generally match the counter metric equivalents when applicable, such as:

nv_inference_request_success{model="my_model",version="1"} 161
nv_inference_queue_duration_us{model="my_model",version="1"} 11110

Triton has a set of default quantiles to track, as shown above. To set custom quantiles, you can use the --metrics-config CLI option. The format is:

tritonserver --metrics-config summary_quantiles="<quantile1>:<error1>,...,<quantileN>:<errorN>"`

For example:

tritonserver --metrics-config summary_quantiles="0.5:0.05,0.9:0.01,0.95:0.001,0.99:0.001"`

To better understand the setting of error values for computing each quantile, see the best practices for histograms and summaries.

GPU Metrics

GPU metrics are collected through the use of DCGM. Collection of GPU metrics can be toggled with the --allow-gpu-metrics CLI flag. If building Triton locally, the TRITON_ENABLE_METRICS_GPU CMake build flag can be used to toggle building the relevant code entirely.

CPU Metrics

Collection of CPU metrics can be toggled with the --allow-cpu-metrics CLI flag. If building Triton locally, the TRITON_ENABLE_METRICS_CPU CMake build flag can be used to toggle building the relevant code entirely.

Note

CPU Metrics are currently only supported on Linux. They collect information from the /proc filesystem such as /proc/stat and /proc/meminfo.

Pinned Memory Metrics

Starting in 24.01, Triton offers Pinned Memory metrics to monitor the utilization of the Pinned Memory pool.

Response Cache Metrics

Cache metrics can be reported in two ways:

1. A base set of cache metrics will be reported by Triton directly, such as the cache hit/miss counts and durations described below.

2. As of 23.03, additional cache metrics may be reported depending on the cache implementation being used through Triton's Metrics API.

Triton-reported Response Cache Metrics

Compute latency metrics in the Inference Request Metrics table above are calculated for the time spent in model inference backends. If the response cache is enabled for a given model (see Response Cache docs for more info), total inference times may be affected by response cache lookup times.

On cache hits, "Cache Hit Time" indicates the time spent looking up the response, and "Compute Input Time" / "Compute Time" / "Compute Output Time" are not recorded.

On cache misses, "Cache Miss Time" indicates the time spent looking up the request hash and inserting the computed output tensor data into the cache. Otherwise, "Compute Input Time" / "Compute Time" / "Compute Output Time" will be recorded as usual.

Similar to the Summaries section above for Inference Request Metrics, the per-model cache hit/miss latency metrics also support Summaries.

Note

For models with response caching enabled, the inference request summary metric is currently disabled. This is due to extra time spent internally on cache management that wouldn't be reflected correctly in the end to end request time. Other summary metrics are unaffected.

Custom Metrics

Triton exposes a C API to allow users and backends to register and collect custom metrics with the existing Triton metrics endpoint. The user takes the ownership of the custom metrics created through the APIs and must manage their lifetime following the API documentation.

The identity_backend demonstrates a practical example of adding a custom metric to a backend.

Further documentation can be found in the TRITONSERVER_MetricFamily* and TRITONSERVER_Metric* API annotations in tritonserver.h.

TensorRT-LLM Backend Metrics

The TRT-LLM backend uses the custom metrics API to track and expose specific metrics about LLMs, KV Cache, and Inflight Batching to Triton: https://github.com/triton-inference-server/tensorrtllm_backend?tab=readme-ov-file#triton-metrics

vLLM Backend Metrics

The vLLM backend uses the custom metrics API to track and expose specific metrics about LLMs to Triton: https://github.com/triton-inference-server/vllm_backend?tab=readme-ov-file#triton-metrics