Reference

AIPerf Metrics Reference

View as Markdown

This document provides a comprehensive reference of all metrics available in AIPerf for benchmarking LLM inference performance. Metrics are organized by computation type to help you understand when and how each metric is calculated.

Table of Contents

Quick Reference

The sections below provide detailed descriptions, requirements, and notes for each metric.

Understanding Metric Types

AIPerf computes metrics in three distinct phases during benchmark execution: Record Metrics, Aggregate Metrics, and Derived Metrics.

Record Metrics

Record Metrics are computed individually for each request and its response(s) during the benchmark run. A single request may have one response (non-streaming) or multiple responses (streaming). These metrics capture per-request characteristics such as latency, token counts, and streaming behavior. Record metrics produce statistical distributions (min, max, mean, median, p90, p99, etc.) that reveal performance variability across requests.

Example Metrics

request_latency, time_to_first_token, inter_token_latency, output_token_count, input_sequence_length

Dependencies

Record Metrics can depend on raw request/response data and other Record Metrics from the same request.

Example Scenario

request_latency measures the time for each individual request from start to final response. If you send 100 requests, you get 100 latency values that form a distribution showing how latency varies across requests.

Aggregate Metrics

Aggregate Metrics are computed by tracking or accumulating values across all requests in real-time during the benchmark. These include counters, min/max timestamps, and other global statistics. Aggregate metrics produce a single value representing the entire benchmark run.

Example Metrics

request_count, error_request_count, min_request_timestamp, max_response_timestamp

Dependencies

Aggregate Metrics can depend on raw request/response data, Record Metrics and other Aggregate Metrics.

Example Scenario

request_count increments by 1 for each successful request. At the end of a benchmark with 100 successful requests, this metric equals 100 (a single value, not a distribution).

Derived Metrics

Derived Metrics are computed by applying mathematical formulas to other metric results, but are not computed per-record like Record Metrics. Instead, these metrics depend on one or more prerequisite metrics being available first and are calculated either after the benchmark completes for final results or in real-time across all current data for live metrics display. Derived metrics can produce either single values or distributions depending on their dependencies.

Example Metrics

request_throughput, output_token_throughput, benchmark_duration

Dependencies

Derived Metrics can depend on Record Metrics, Aggregate Metrics, and other Derived Metrics, but do not have any knowledge of the individual request/response data.

Example Scenario

request_throughput is computed from request_count / benchmark_duration_seconds. This requires both request_count and benchmark_duration to be available first, then applies a formula to produce a single throughput value (e.g., 10.5 requests/sec).

Detailed Metric Descriptions

Streaming Metrics

All metrics in this section require the --streaming flag with a token-producing endpoint and at least one non-empty response chunk.

Time to First Token (TTFT)

Type: Record Metric

Measures how long it takes to receive the first token (or chunk of tokens) after sending a request. This is critical for user-perceived responsiveness in streaming scenarios, as it represents how quickly the model begins generating output.

Formula:

1# nanoseconds
2ttft_ns = request.content_responses[0].perf_ns - request.start_perf_ns
3
4# Convert to milliseconds for display
5ttft_ms = ttft_ns / 1e6
6
7# Convert to seconds for throughput calculations
8ttft_seconds = ttft_ns / 1e9

Notes:

  • Includes network latency, queuing time, prompt processing, and generation of the first token (or chunk of tokens).
  • Raw timestamps are in nanoseconds; converted to milliseconds for display and seconds for rate calculations.
  • Response chunks refer to individual messages with non-empty content received during streaming.

Time to Second Token (TTST)

Type: Record Metric

Measures the time gap between the first and second chunk of tokens. This metric helps identify generation startup overhead separate from steady-state streaming throughput.

Formula:

1# nanoseconds
2ttst_ns = request.content_responses[1].perf_ns - request.content_responses[0].perf_ns
3
4# Convert to milliseconds for display
5ttst_ms = ttst_ns / 1e6

Notes:

  • Requires at least 2 non-empty response chunks to compute the time between first and second tokens.
  • Raw timestamps are in nanoseconds; converted to milliseconds for display.

Time to First Output Token (TTFO)

Type: Record Metric

Calculates the time elapsed from request start to the first non-reasoning output token. This metric measures the latency from when a request is initiated to when the first actual output token (non-reasoning content) is received. It is particularly relevant for models that perform extended reasoning before generating output.

Formula:

1# nanoseconds
2# First non-reasoning token: TextResponseData with non-empty text, or
3# ReasoningResponseData with non-empty content field
4ttfo_ns = first_non_reasoning_token_perf_ns - request.start_perf_ns
5
6# Convert to milliseconds for display
7ttfo_ms = ttfo_ns / 1e6

Notes:

  • TTFO vs TTFT: Time to First Output (TTFO) measures time to the first non-reasoning token, while Time to First Token (TTFT) measures time to any first token including reasoning tokens. For models without reasoning, TTFO and TTFT are equivalent.
  • Non-reasoning tokens include TextResponseData with non-empty text, or ReasoningResponseData with non-empty content field (regardless of reasoning field).
  • Requires at least one non-empty non-reasoning response chunk.

Inter Token Latency (ITL)

Type: Record Metric

Measures the average time between consecutive tokens during generation, excluding the initial TTFT overhead. This represents the steady-state token generation rate.

Formula:

1# Calculate in nanoseconds, then convert to seconds
2inter_token_latency_ns = (request_latency_ns - time_to_first_token_ns) / (output_sequence_length - 1)
3
4# Convert to seconds for throughput calculations
5inter_token_latency_seconds = inter_token_latency_ns / 1e9
6
7# Convert to milliseconds for display
8inter_token_latency_ms = inter_token_latency_ns / 1e6

Notes:

  • Requires at least 2 non-empty response chunks and valid time_to_first_token, request_latency, and output_sequence_length metrics.
  • Result is in seconds when used for throughput calculations (Output Token Throughput Per User).

Inter Chunk Latency (ICL)

Type: Record Metric

Captures the time gaps between all consecutive response chunks in a streaming response, providing a distribution of chunk arrival times rather than a single average. Note that this is different from the ITL metric, which measures the time between consecutive tokens regardless of chunk size.

Formula:

1inter_chunk_latency = [request.content_responses[i].perf_ns - request.content_responses[i-1].perf_ns for i in range(1, len(request.content_responses))]

Notes:

  • Requires at least 2 response chunks.
  • Unlike ITL (which produces a single average), ICL provides the full distribution of inter-chunk times.
  • Useful for detecting variability, jitter, or issues in streaming delivery.
  • Analyzing ICL distributions can reveal batching behavior, scheduling issues, or network variability.

Output Token Throughput Per User

Type: Record Metric

This metric is computed per-request, and it excludes the TTFT from the equation, so it is not directly comparable to the Output Token Throughput metric.

The token generation rate experienced by an individual user/request, measured as the inverse of inter-token latency. This represents single-request streaming performance.

Formula:

1output_token_throughput_per_user = 1.0 / inter_token_latency_seconds

Notes:

  • Computes the inverse of ITL to show tokens per second from an individual user’s perspective.
  • Differs from Output Token Throughput (aggregate across all concurrent requests) by focusing on single-request experience.
  • Useful for understanding the user experience independent of concurrency effects.

Prefill Throughput Per User

Type: Record Metric

Measures the rate at which input tokens are processed during the prefill phase, calculated as input tokens per second based on TTFT. This is only applicable to streaming responses.

Formula:

1prefill_throughput_per_user = input_sequence_length / time_to_first_token_seconds

Notes:

  • Higher values indicate faster prompt processing.
  • Useful for understanding input processing capacity and bottlenecks.
  • Depends on Input Sequence Length and TTFT metrics.

Token Based Metrics

All metrics in this section require token-producing endpoints that return text content (chat, completion, etc.). These metrics are not available for embeddings or other non-generative endpoints.

Output Token Count

Type: Record Metric

The number of output tokens generated for a single request, excluding reasoning tokens. This represents the output tokens returned to the user across all responses for the request.

Formula:

1output_token_count = len(tokenizer.encode(content, add_special_tokens=False))

Notes:

  • Tokenization uses add_special_tokens=False to count only content tokens, excluding special tokens added by the tokenizer.
  • For streaming requests with multiple responses, the responses are joined together and then tokens are counted.
  • For models that expose reasoning in a separate reasoning_content field, this metric counts only non-reasoning output tokens.
  • If reasoning appears inside the regular content (e.g., <think> blocks), those tokens will be counted unless explicitly filtered.

Output Sequence Length (OSL)

Type: Record Metric

The total number of completion tokens (output + reasoning) generated for a single request across all its responses. This represents the complete token generation workload for the request.

Formula:

1output_sequence_length = (output_token_count or 0) + (reasoning_token_count or 0)

Notes:

  • For models that do not support/separate reasoning tokens, OSL equals the output token count.

Input Sequence Length (ISL)

Type: Record Metric

The number of input/prompt tokens for a single request. This represents the size of the input sent to the model.

Formula:

1input_sequence_length = len(tokenizer.encode(prompt, add_special_tokens=False))

Notes:

  • Tokenization uses add_special_tokens=False to count only content tokens, excluding special tokens added by the tokenizer.
  • Useful for understanding the relationship between input size and latency/throughput.

Total Output Tokens

Type: Derived Metric

The sum of all output tokens (excluding reasoning tokens) generated across all requests. This represents the total output token workload.

Formula:

1total_output_tokens = sum(r.output_token_count for r in records if r.valid)

Notes:

  • Aggregates output tokens across all successful requests.
  • Useful for capacity planning and cost estimation.

Total Output Sequence Length

Type: Derived Metric

The sum of all completion tokens (output + reasoning) generated across all requests. This represents the complete token generation workload.

Formula:

1total_osl = sum(r.output_sequence_length for r in records if r.valid)

Notes:

  • Aggregates the complete token generation workload including both output and reasoning tokens.
  • For models without reasoning tokens, this equals Total Output Tokens.

Total Input Sequence Length

Type: Derived Metric

The sum of all input/prompt tokens processed across all requests. This represents the total input workload sent to the model.

Formula:

1total_isl = sum(r.input_sequence_length for r in records if r.valid)

Notes:

  • Useful for understanding the input workload, capacity planning, and analyzing the relationship between input size and system performance.

E2E Output Token Throughput

Type: Record Metric

Per-request output token throughput based on end-to-end request latency. Unlike Output Token Throughput Per User (which uses 1/ITL and excludes TTFT), this metric includes TTFT, queuing, and all other overhead in the denominator. Available for both streaming and non-streaming responses.

Formula:

1e2e_output_token_throughput = output_sequence_length / request_latency_seconds

Notes:

  • Uses total request latency (not ITL), so values will be slightly lower than Output Token Throughput Per User for streaming responses.
  • Available for non-streaming responses (unlike Output Token Throughput Per User which requires streaming).
  • Flags: PRODUCES_TOKENS_ONLY | LARGER_IS_BETTER
  • Depends on Output Sequence Length and Request Latency metrics.

Output Token Throughput

Type: Derived Metric

This metric is computed as a single value across all requests and includes TTFT in the equation, so it is not directly comparable to the Output Token Throughput Per User metric.

The aggregate token generation rate across all concurrent requests, measured as total tokens per second. This represents the system’s overall token generation capacity.

Formula:

1output_token_throughput = total_osl / benchmark_duration_seconds

Notes:

  • Measures aggregate throughput across all concurrent requests; represents the overall system token generation rate.
  • Higher values indicate better system utilization and capacity.

Total Token Throughput

Type: Derived Metric

Calculates the total token throughput metric, combining both input and output token processing across all concurrent requests.

Formula:

1total_token_throughput = (total_isl + total_osl) / benchmark_duration_seconds

Notes:

  • Measures the combined input and output token processing rate.
  • Includes reasoning tokens in the output count (via total_osl).
  • Useful for understanding total system token processing capacity.

Image Metrics

All metrics in this section require image-capable endpoints (e.g., image generation APIs). These metrics are not available for text-only or other non-image endpoints.

Number of Images

Type: Record Metric

The number of images in the request, summed across all turns. This is the foundation metric used by Image Throughput and Image Latency.

Formula:

1num_images = sum(len(image.contents) for turn in request.turns for image in turn.images)

Notes:

  • Requires at least one image in at least one turn.
  • Not displayed in console output (NO_CONSOLE flag).

Image Throughput

Type: Record Metric

Calculates the image throughput from the record by dividing the number of images by the request latency.

Formula:

1image_throughput = num_images / request_latency_seconds

Notes:

  • Higher values indicate faster image generation.

Image Latency

Type: Record Metric

Calculates the image latency from the record by dividing the request latency by the number of images.

Formula:

1image_latency = request_latency_ms / num_images

Notes:

  • Lower values indicate faster per-image generation.

Video Metrics

All metrics in this section require video-producing endpoints (e.g., SGLang video generation). These metrics rely on server-reported fields in the response and are not available for non-video endpoints.

Video Inference Time

Type: Record Metric

Server-reported GPU generation time for video inference, extracted from the inference_time_s field in video generation responses (e.g., SGLang).

Formula:

1video_inference_time = response.data.inference_time_s

Notes:

  • Value comes from the server, not computed by AIPerf.
  • Displayed in milliseconds.

Video Peak Memory

Type: Record Metric

Server-reported peak GPU memory usage during video generation, extracted from the peak_memory_mb field in video generation responses.

Formula:

1video_peak_memory = response.data.peak_memory_mb

Notes:

  • Value comes from the server, not computed by AIPerf.
  • Unit is megabytes.

Reasoning Metrics

All metrics in this section require models and backends that expose reasoning content in a separate reasoning_content field, distinct from the regular content field.

Reasoning Token Count

Type: Record Metric

The number of reasoning tokens generated for a single request. These are tokens used for “thinking” or chain-of-thought reasoning before generating the final output.

Formula:

1reasoning_token_count = len(tokenizer.encode(reasoning_content, add_special_tokens=False))

Notes:

  • Tokenization uses add_special_tokens=False to count only content tokens, excluding special tokens added by the tokenizer.
  • Does not differentiate <think> tags or extract reasoning from within the regular content field.

Total Reasoning Tokens

Type: Derived Metric

The sum of all reasoning tokens generated across all requests. This represents the total reasoning/thinking workload.

Formula:

1total_reasoning_tokens = sum(r.reasoning_token_count for r in records if r.valid)

Notes:

  • Useful for understanding the reasoning overhead and cost for reasoning-enabled models.

Usage Field Metrics

All metrics in this section track API-reported token counts from the usage field in API responses. These are not displayed in console output but are available in exports. These metrics are useful for comparing client-side token counts with server-reported counts to detect discrepancies.

Usage Prompt Tokens

Type: Record Metric

The number of input/prompt tokens as reported by the API’s usage.prompt_tokens field for a single request.

Formula:

1usage_prompt_tokens = response.usage.prompt_tokens  # from last non-None response

Notes:

  • Taken from the API response usage object, not computed by AIPerf.
  • May differ from client-side Input Sequence Length due to different tokenizers or special tokens.
  • For streaming responses, uses the last non-None value reported.

Usage Completion Tokens

Type: Record Metric

The number of completion tokens as reported by the API’s usage.completion_tokens field for a single request.

Formula:

1usage_completion_tokens = response.usage.completion_tokens  # from last non-None response

Notes:

  • Taken from the API response usage object, not computed by AIPerf.
  • May differ from client-side Output Sequence Length due to different tokenizers or counting methods.
  • For streaming responses, uses the last non-None value reported.

Usage Total Tokens

Type: Record Metric

The total number of tokens (prompt + completion) as reported by the API’s usage.total_tokens field for a single request.

Formula:

1usage_total_tokens = response.usage.total_tokens  # from last non-None response

Notes:

  • Taken from the API response usage object, not computed by AIPerf.
  • Should generally equal usage_prompt_tokens + usage_completion_tokens.
  • For streaming responses, uses the last non-None value reported.

Usage Reasoning Tokens

Type: Record Metric

The number of reasoning tokens as reported by the API’s usage.completion_tokens_details.reasoning_tokens field for a single request. Only available for reasoning-enabled models.

Formula:

1usage_reasoning_tokens = response.usage.completion_tokens_details.reasoning_tokens

Notes:

  • Taken from the API response for reasoning-enabled models.
  • May differ from client-side Reasoning Token Count due to different tokenizers.
  • For streaming responses, uses the last non-None value reported.

Total Usage Prompt Tokens

Type: Derived Metric

The sum of all API-reported prompt tokens across all requests.

Formula:

1total_usage_prompt_tokens = sum(r.usage_prompt_tokens for r in records if r.valid)

Notes:

  • Aggregates server-reported input tokens across all requests.

Total Usage Completion Tokens

Type: Derived Metric

The sum of all API-reported completion tokens across all requests.

Formula:

1total_usage_completion_tokens = sum(r.usage_completion_tokens for r in records if r.valid)

Notes:

  • Aggregates server-reported completion tokens across all requests.

Total Usage Total Tokens

Type: Derived Metric

The sum of all API-reported total tokens across all requests.

Formula:

1total_usage_total_tokens = sum(r.usage_total_tokens for r in records if r.valid)

Notes:

  • Aggregates server-reported total tokens across all requests.

Usage Discrepancy Metrics

These metrics measure the percentage difference between API-reported token counts (usage fields) and client-computed token counts. They are not displayed in console output but help identify tokenizer mismatches or counting discrepancies.

Usage Prompt Tokens Diff %

Type: Record Metric

The percentage difference between API-reported prompt tokens and client-computed Input Sequence Length.

Formula:

1usage_prompt_tokens_diff_pct = abs((usage_prompt_tokens - input_sequence_length) / input_sequence_length) * 100

Notes:

  • Values close to 0% indicate good agreement between client and server token counts.
  • Large differences may indicate tokenizer mismatches or special token handling differences.

Usage Completion Tokens Diff %

Type: Record Metric

The percentage difference between API-reported completion tokens and client-computed Output Sequence Length.

Formula:

1usage_completion_tokens_diff_pct = abs((usage_completion_tokens - output_sequence_length) / output_sequence_length) * 100

Notes:

  • Values close to 0% indicate good agreement between client and server token counts.
  • Large differences may indicate tokenizer mismatches or different counting methods.

Usage Reasoning Tokens Diff %

Type: Record Metric

The percentage difference between API-reported reasoning tokens and client-computed Reasoning Token Count.

Formula:

1usage_reasoning_tokens_diff_pct = abs((usage_reasoning_tokens - reasoning_token_count) / reasoning_token_count) * 100

Notes:

  • Only available for reasoning-enabled models.
  • Values close to 0% indicate good agreement between client and server reasoning token counts.

Usage Discrepancy Count

Type: Aggregate Metric

The number of requests where token count differences exceed a threshold (default 10%).

Formula:

1usage_discrepancy_count = sum(1 for r in records if r.any_diff > threshold)

Notes:

  • Default threshold is 10% difference.
  • Counts requests where prompt, completion, or reasoning token differences are significant.
  • Useful for monitoring overall token count agreement quality.

OSL Mismatch Metrics

These metrics measure the difference between requested output sequence length (--osl/max_tokens) and actual output tokens generated. They help identify when the server is not honoring the requested output length, typically because EOS tokens stop generation early. These metrics are not displayed in console output but are available in exports and used by the end-of-benchmark warning.

OSL Mismatch Diff %

Type: Record Metric

The signed percentage difference between actual output sequence length and requested OSL. Negative values mean the server stopped early (actual < requested), positive values mean it generated more than requested.

Formula:

1osl_mismatch_diff_pct = ((output_sequence_length - requested_osl) / requested_osl) * 100

Notes:

  • Negative = stopped early (hit EOS before max_tokens)
  • Positive = generated more than requested
  • 0% = exact match between requested and actual
  • Example: Requested 100 tokens, got 50 → Diff = -50%
  • Example: Requested 100 tokens, got 120 → Diff = 20%

OSL Mismatch Count

Type: Aggregate Metric

The count of requests where the absolute token difference exceeds the effective threshold. Used to trigger the end-of-benchmark warning panel.

Formula:

1# Effective threshold is capped to be tighter for large OSL values
2threshold_tokens = min(requested_osl * (pct_threshold / 100), max_token_threshold)
3diff_tokens = abs(actual_osl - requested_osl)
4osl_mismatch_count = sum(1 for r in records if diff_tokens > threshold_tokens)

Notes:

  • Default percentage threshold is 5% (AIPERF_METRICS_OSL_MISMATCH_PCT_THRESHOLD).
  • Default max token threshold is 50 (AIPERF_METRICS_OSL_MISMATCH_MAX_TOKEN_THRESHOLD).
  • The min() makes threshold tighter for large OSL: requesting 2000 tokens caps at 50 token diff instead of 100 (5%).
  • Counts both early stops (negative diff) and over-generation (positive diff).
  • When this count is non-zero, a warning panel is displayed at the end of the benchmark.
  • To ensure servers honor --osl, use --extra-inputs ignore_eos:true or --extra-inputs min_tokens:<value>.
  • If discrepancy is due to tokenizer mismatch between client and server, use --use-server-token-count.

Server support for min_tokens:

ServerParameterNotes
vLLMmin_tokensDefault: 0
TensorRT-LLMmin_tokensDefault: 1
SGLangmin_new_tokensDefault: 0
TGImin_new_tokens⚠️ Unclear API support; TGI in maintenance mode

Goodput Metrics

Goodput metrics measure the throughput of requests that meet user-defined Service Level Objectives (SLOs). See the Goodput tutorial for configuration details.

Good Request Count

Type: Aggregate Metric

The number of requests that meet all user-defined SLO thresholds during the benchmark.

Formula:

1good_request_count = sum(1 for r in records if r.all_slos_met)

Notes:

  • Requires SLO thresholds to be configured (e.g., --goodput).
  • Only counts requests where ALL SLO constraints are satisfied.
  • Used to calculate Goodput metric.

Goodput

Type: Derived Metric

The rate of SLO-compliant requests per second. This represents the effective throughput of requests meeting quality requirements.

Formula:

1goodput = good_request_count / benchmark_duration_seconds

Notes:

  • Requires SLO thresholds to be configured.
  • Always less than or equal to Request Throughput.
  • Useful for capacity planning and comparing systems based on quality-adjusted throughput.

Error Metrics

These metrics are computed only for failed/error requests and are not displayed in console output.

Error Input Sequence Length

Type: Record Metric

The number of input tokens for requests that resulted in errors. This helps analyze whether input size correlates with errors.

Formula:

1error_isl = input_sequence_length  # for error requests only

Notes:

  • Only computed for requests that failed.
  • Useful for identifying if certain input sizes trigger errors.

Total Error Input Sequence Length

Type: Derived Metric

The sum of all input tokens from requests that resulted in errors.

Formula:

1total_error_isl = sum(r.input_sequence_length for r in records if not r.valid)

Notes:

  • Aggregates input tokens across all failed requests.

General Metrics

Metrics in this section are available for all benchmark runs with no special requirements.

Request Latency

Type: Record Metric

Measures the total end-to-end time from sending a request until receiving the final response. For streaming requests with multiple responses, this measures until the last response is received. This is the complete time experienced by the client for a single request.

Formula:

1request_latency_ns = request.content_responses[-1].perf_ns - request.start_perf_ns

Notes:

  • Includes all components: network time, queuing, prompt processing, token generation, and response transmission.
  • For streaming requests, measures from request start to the final chunk received.

Request Throughput

Type: Derived Metric

The overall rate of completed requests per second across the entire benchmark. This represents the system’s ability to process requests under the given concurrency and load.

Formula:

1request_throughput = request_count / benchmark_duration_seconds

Notes:

  • Captures the aggregate request processing rate; higher values indicate better system throughput.
  • Affected by concurrency level, request complexity, output sequence length, and system capacity.

Request Count

Type: Aggregate Metric

The total number of successfully completed requests in the benchmark. This includes all requests that received valid responses, regardless of streaming mode.

Formula:

1request_count = sum(1 for r in records if r.valid)

Error Request Count

Type: Aggregate Metric

The total number of failed/error requests encountered during the benchmark. This includes network errors, HTTP errors, timeout errors, and other failures.

Formula:

1error_request_count = sum(1 for r in records if not r.valid)

Notes:

  • Error rate can be computed as error_request_count / (request_count + error_request_count).

Minimum Request Timestamp

Type: Aggregate Metric

The wall-clock timestamp of the first request sent in the benchmark. This is used to calculate the benchmark duration and represents the start of the benchmark run.

Formula:

1min_request_timestamp = min(r.timestamp_ns for r in records)

Maximum Response Timestamp

Type: Aggregate Metric

The wall-clock timestamp of the last response received in the benchmark. This is used to calculate the benchmark duration and represents the end of the benchmark run.

Formula:

1max_response_timestamp = max(r.timestamp_ns + r.request_latency for r in records)

Benchmark Duration

Type: Derived Metric

The total elapsed time from the first request sent to the last response received. This represents the complete wall-clock duration of the benchmark run.

Formula:

1benchmark_duration = max_response_timestamp - min_request_timestamp

Notes:

  • Uses wall-clock timestamps representing real calendar time.
  • Used as the denominator for throughput calculations; represents the effective measurement window.

HTTP Trace Metrics

All metrics in this section require HTTP trace data to be collected during requests. These metrics provide detailed HTTP request lifecycle timing following k6 naming conventions. See the HTTP Trace Metrics tutorial for configuration details.

HTTP Blocked

Type: Record Metric

Time spent blocked waiting for a free TCP connection slot from the pool. This metric measures the time a request spent waiting in the connection pool queue before a connection became available. High values indicate connection pool saturation.

Formula:

1http_req_blocked = connection_pool_wait_end_perf_ns - connection_pool_wait_start_perf_ns

Notes:

  • k6 equivalent: http_req_blocked
  • HAR equivalent: blocked
  • Returns 0 if no pool wait occurred (connection immediately available).
  • Only available for AioHttpTraceData.

HTTP DNS Lookup

Type: Record Metric

Time spent on DNS resolution. This metric measures the time spent resolving the hostname to an IP address.

Formula:

1http_req_dns_lookup = dns_lookup_end_perf_ns - dns_lookup_start_perf_ns

Notes:

  • k6 equivalent: http_req_looking_up
  • HAR equivalent: dns
  • Returns 0 if DNS cache hit or connection was reused.
  • Only available for AioHttpTraceData.

HTTP Connecting

Type: Record Metric

Time spent establishing TCP connection to the remote host. For HTTPS requests, this includes both TCP connection establishment and TLS handshake time (combined measurement from aiohttp).

Formula:

1http_req_connecting = tcp_connect_end_perf_ns - tcp_connect_start_perf_ns

Notes:

  • k6 equivalent: http_req_connecting
  • HAR equivalent: connect
  • Returns 0 if connection was reused.
  • Only available for AioHttpTraceData.

HTTP Sending

Type: Record Metric

Time spent sending data to the remote host. This metric measures the time from when the request started being sent to when the full request (headers + body) was transmitted.

Formula:

1http_req_sending = request_send_end_perf_ns - request_send_start_perf_ns

Notes:

  • k6 equivalent: http_req_sending
  • HAR equivalent: send

HTTP Waiting (TTFB)

Type: Record Metric

Time to First Byte (TTFB) - time waiting for the server to respond. This metric measures the time from when the request was fully sent to when the first byte of the response body was received. This represents server processing time plus network latency.

Formula:

1http_req_waiting = response_receive_start_perf_ns - request_send_end_perf_ns

Notes:

  • k6 equivalent: http_req_waiting (also known as TTFB)
  • HAR equivalent: wait
  • Note that this is not the same as the time to first token (TTFT), which is the time from request start to the first valid token received. The server may send non-token data first.

HTTP Receiving

Type: Record Metric

Time spent receiving response data from the remote host. This metric measures the time from when the first byte of the response was received to when the last byte was received.

Formula:

1http_req_receiving = response_receive_end_perf_ns - response_receive_start_perf_ns

Notes:

  • k6 equivalent: http_req_receiving
  • HAR equivalent: receive
  • Returns 0 if response was a single chunk.

HTTP Duration

Type: Record Metric

Time for HTTP request/response exchange, excluding connection overhead. This measures only the request/response exchange time: sending + waiting + receiving.

Formula:

1http_req_duration = response_receive_end_perf_ns - request_send_start_perf_ns

Notes:

  • k6 equivalent: http_req_duration
  • HAR equivalent: time
  • EXCLUDES connection overhead (blocked, dns_lookup, connecting).
  • For full end-to-end time including connection setup, use http_req_total.
  • Note: This uses trace-level timestamps for more accurate measurement than application-level request latency.

HTTP Connection Overhead

Type: Record Metric

Total connection overhead time (blocked + dns_lookup + connecting). This metric combines all pre-request overhead.

Formula:

1http_req_connection_overhead = http_req_blocked + http_req_dns_lookup + http_req_connecting

Notes:

  • Useful for identifying total connection establishment costs.
  • Returns 0 if connection was reused with no pool wait.
  • Only available for AioHttpTraceData.

HTTP Total Time

Type: Record Metric

Sum of all HTTP timing phases from connection pool to last chunk received. This is the sum of all 6 timing components: blocked + dns_lookup + connecting + sending + waiting + receiving.

Formula:

1http_req_total = http_req_blocked + http_req_dns_lookup + http_req_connecting + http_req_sending + http_req_waiting + http_req_receiving

Notes:

  • This ensures the math adds up: individual timing metrics sum exactly to this total.
  • Only available for AioHttpTraceData (requires connection overhead metrics).

HTTP Data Sent

Type: Record Metric

Total bytes sent in the HTTP request (headers + body).

Formula:

1http_req_data_sent = trace.request_bytes_total

Notes:

  • k6 equivalent: data_sent (per request)
  • Measures total bytes written to the transport layer.

HTTP Data Received

Type: Record Metric

Total bytes received in the HTTP response (headers + body).

Formula:

1http_req_data_received = trace.response_bytes_total

Notes:

  • k6 equivalent: data_received (per request)
  • Measures total bytes read from the transport layer.

HTTP Connection Reused

Type: Record Metric

Whether the HTTP connection was reused from the connection pool. Returns 1 if reused, 0 if new connection was established.

Formula:

1http_req_connection_reused = 1 if connection_reused_perf_ns is not None else 0

Notes:

  • Helps identify connection reuse patterns and keep-alive effectiveness.
  • Only available for AioHttpTraceData.

HTTP Chunks Sent

Type: Record Metric

Number of transport-level write operations during the request. Useful for debugging chunked transfers.

Formula:

1http_req_chunks_sent = trace.request_chunks_count

Notes:

  • Not displayed in console output (NO_CONSOLE flag).

HTTP Chunks Received

Type: Record Metric

Number of transport-level read operations during the response. Useful for debugging chunked/streaming responses.

Formula:

1http_req_chunks_received = trace.response_chunks_count

Notes:

  • Not displayed in console output (NO_CONSOLE flag).

Multi-Run Aggregate Metrics

These metrics are only available when using --num-profile-runs > 1 for confidence reporting.

When running multiple profile iterations with --num-profile-runs, AIPerf computes aggregate statistics across all runs to quantify measurement variance and repeatability. These statistics are written to aggregate/profile_export_aiperf_aggregate.json and aggregate/profile_export_aiperf_aggregate.csv.

For detailed information about aggregate statistics, their mathematical definitions, and interpretation guidelines, see the Multi-Run Confidence Tutorial.

Quick Reference

The following aggregate statistics are computed for each metric:

  • mean: Average value across all runs
  • std: Standard deviation (measure of spread)
  • min: Minimum value observed
  • max: Maximum value observed
  • cv: Coefficient of Variation (normalized variability)
  • se: Standard Error (uncertainty in the mean)
  • ci_low, ci_high: Confidence interval bounds
  • t_critical: t-distribution critical value used

Aggregate Metadata

The aggregate output also includes metadata about the multi-run benchmark:

  • aggregation_type: Always “confidence” for multi-run confidence reporting
  • num_profile_runs: Total number of runs requested
  • num_successful_runs: Number of runs that completed successfully
  • failed_runs: List of failed runs with error details
  • confidence_level: Confidence level used for intervals (e.g., 0.95)
  • cooldown_seconds: Cooldown duration between runs
  • run_labels: Labels for each run (e.g., [“run_0001”, “run_0002”, …])

Metric Flags Reference

Metric flags are used to control when and how metrics are computed, displayed, and grouped. Flags can be combined using bitwise operations to create composite behaviors.

Individual Flags

FlagDescriptionImpact
NONENo flags setMetric has default behavior with no special restrictions
STREAMING_ONLYOnly computed for streaming responsesRequires Server-Sent Events (SSE) with multiple response chunks; skipped for non-streaming requests
ERROR_ONLYOnly computed for error requestsTracks error-specific information; computed only for invalid/failed requests
PRODUCES_TOKENS_ONLYOnly computed for token-producing endpointsRequires endpoints that return text/token content; skipped for embeddings and non-generative endpoints
NO_CONSOLENot displayed in console outputMetric computed but excluded from terminal display; available in JSON/CSV/JSONL exports and used by other metrics
LARGER_IS_BETTERHigher values indicate better performanceUsed for throughput and count metrics to indicate optimization direction
INTERNALInternal AIPerf metricUsed for AIPerf system diagnostics; not displayed in console or exported without developer mode
SUPPORTS_AUDIO_ONLYOnly computed for audio endpointsRequires audio-capable endpoints; skipped for other endpoint types
SUPPORTS_IMAGE_ONLYOnly computed for image endpointsRequires image-capable endpoints; skipped for other endpoint types
SUPPORTS_REASONINGRequires reasoning token supportOnly available for models and endpoints that expose reasoning content in separate fields
EXPERIMENTALExperimental/unstable metricMay change or be removed in future releases; not displayed in console or exported without developer mode
GOODPUTOnly computed when goodput is enabledRequires SLO thresholds to be configured (e.g., --goodput); skipped otherwise
NO_INDIVIDUAL_RECORDSNot exported for individual recordsAggregate metrics not relevant to individual records (e.g., request count, min/max timestamps); excluded from per-record exports
TOKENIZES_INPUT_ONLYOnly computed when endpoint tokenizes inputRequires endpoints that process and tokenize input text; skipped for non-text endpoints
HTTP_TRACE_ONLYOnly computed when HTTP trace data is availableRequires HTTP request tracing to be enabled; provides detailed HTTP lifecycle timing metrics
SUPPORTS_VIDEO_ONLYOnly computed for video endpointsRequires video-capable endpoints; skipped for other endpoint types
USAGE_DIFF_ONLYOnly computed when usage field data is availableRequires API responses to include usage field with token counts for comparison with client-computed values
PRODUCES_VIDEO_ONLYOnly computed for video-producing endpointsRequires endpoints that produce video output (e.g., SGLang video generation)

Composite Flags

These flags are combinations of multiple individual flags for convenience:

FlagCompositionDescription
STREAMING_TOKENS_ONLYSTREAMING_ONLY + PRODUCES_TOKENS_ONLYRequires both streaming support and token-producing endpoints