Skip to content

arrival-patterns.md

Latest commit

 

History

History
300 lines (230 loc) · 9.51 KB

arrival-patterns.md

File metadata and controls

300 lines (230 loc) · 9.51 KB

Arrival Patterns: Simulating Realistic Traffic

When benchmarking with --request-rate, AIPerf can vary how requests arrive over time. The --arrival-pattern option controls the distribution of inter-arrival times, letting you simulate everything from perfectly regular traffic to bursty real-world patterns.

Why Arrival Patterns Matter

Real traffic doesn't arrive at perfectly regular intervals. Traffic comes in bursts—quiet periods followed by sudden spikes. How your server handles this variance affects real-world performance.

Constant Pattern:         Poisson Pattern:         Gamma (bursty):
|  |  |  |  |  |  |       |   | || |    | |         |||    |    |||  |
└──────────────────▶     └──────────────────▶    └──────────────────▶
  Perfect spacing          Natural variance        Clustered bursts
  (unrealistic)            (typical traffic)       (stress testing)

Quick Start

# Default: Poisson (realistic)
aiperf profile --request-rate 50 ...

# Explicit: Constant (deterministic)
aiperf profile --request-rate 50 --arrival-pattern constant ...

# Bursty: Gamma with low smoothness
aiperf profile --request-rate 50 --arrival-pattern gamma --arrival-smoothness 0.5 ...

Available Patterns

Constant

--arrival-pattern constant

Requests arrive at perfectly regular intervals: exactly 1/rate seconds apart.

Inter-arrival times:
10 QPS → every 100ms:  |····|····|····|····|····|····|
                       0   100  200  300  400  500  600 ms

Use cases:

  • Baseline measurements with no variance
  • Debugging timing issues
  • Comparing against variable patterns
  • Deterministic, reproducible tests

Poisson (Default)

--arrival-pattern poisson

Requests arrive according to a Poisson process—the mathematical model for random events at a constant average rate. Inter-arrival times follow an exponential distribution.

Inter-arrival times (exponential):
10 QPS average:  |··|······|·|···|····|··|·······|···|
                 Varied gaps, same average rate over time

Characteristics:

  • Mean inter-arrival = 1/rate (same as constant)
  • Variance = (1/rate)² (natural randomness)
  • Sometimes requests cluster, sometimes gaps appear
  • Models real user behavior where arrivals are independent

Use cases:

  • Default realistic traffic simulation
  • Standard load testing
  • Comparing to theoretical queueing models

Gamma (Tunable Burstiness)

--arrival-pattern gamma --arrival-smoothness <value>

Gamma distribution generalizes Poisson with a smoothness parameter that controls how bursty or regular arrivals are:

Smoothness Behavior Variance Use Case
< 1.0 Bursty — clustered arrivals with gaps Higher Stress testing, worst-case scenarios
= 1.0 Poisson — natural randomness Medium Same as --arrival-pattern poisson
> 1.0 Smooth — more regular arrivals Lower Controlled testing, less noise 
Smoothness = 0.5 (bursty):
||||      |||        |||||    ||
 Clusters of requests with quiet gaps

Smoothness = 1.0 (Poisson):
|  || |   | |  ||  |   | ||  |
 Natural variance

Smoothness = 2.0 (smooth):
| | | |  | | | | | |  | | | |
 More regular, approaches constant

Mathematical note: The smoothness parameter is the Gamma distribution's shape parameter (k). Scale is automatically computed to maintain the correct mean rate.

Concurrency Burst

# No --request-rate, just --concurrency
aiperf profile --concurrency 50 ...

When you omit --request-rate and only specify --concurrency, AIPerf uses burst mode: zero delay between request dispatches, limited only by the concurrency semaphore.

Burst mode (concurrency=3):
[Req1]────────────────────────────▶
[Req2]────────────────────────────▶
[Req3]────────────────────────────▶
      [Req4]──────────────────────▶  ← Starts when any slot frees

Use cases:

  • Maximum throughput discovery
  • Saturation testing
  • Finding server capacity limits

vLLM Compatibility

AIPerf's --arrival-smoothness is compatible with vLLM's --burstiness parameter:

# Same distribution as vLLM with --burstiness 0.5
aiperf profile \
    --request-rate 50 \
    --arrival-pattern gamma \
    --arrival-smoothness 0.5 \
    ...

This allows direct comparison between AIPerf and vLLM benchmark results when using the same smoothness/burstiness value.

Examples

Baseline vs Realistic Comparison

Compare how your server handles ideal vs realistic traffic:

# Run 1: Constant (baseline)
aiperf profile \
    --model your-model \
    --url localhost:8000 \
    --endpoint-type chat \
    --streaming \
    --request-rate 100 \
    --arrival-pattern constant \
    --benchmark-duration 60 \
    --output-dir results/constant

# Run 2: Poisson (realistic)
aiperf profile \
    --model your-model \
    --url localhost:8000 \
    --endpoint-type chat \
    --streaming \
    --request-rate 100 \
    --arrival-pattern poisson \
    --benchmark-duration 60 \
    --output-dir results/poisson

Compare TTFT and throughput between runs. Higher variance under Poisson indicates sensitivity to traffic patterns.

Stress Testing with Bursty Traffic

Test how your server handles request bursts:

aiperf profile \
    --model your-model \
    --url localhost:8000 \
    --endpoint-type chat \
    --streaming \
    --request-rate 100 \
    --arrival-pattern gamma \
    --arrival-smoothness 0.3 \
    --benchmark-duration 120

Smoothness of 0.3 creates highly bursty traffic—several requests arrive nearly simultaneously, then quiet periods.

Smooth Traffic for Noise Reduction

Reduce variance in measurements for controlled experiments:

aiperf profile \
    --model your-model \
    --url localhost:8000 \
    --endpoint-type chat \
    --streaming \
    --request-rate 50 \
    --arrival-pattern gamma \
    --arrival-smoothness 5.0 \
    --benchmark-duration 60

Smoothness of 5.0 produces very regular arrivals, reducing measurement noise while still having some natural variance.

Progressive Burstiness Test

Run multiple benchmarks with increasing burstiness to find where performance degrades:

for smoothness in 2.0 1.0 0.7 0.5 0.3; do
    aiperf profile \
        --model your-model \
        --url localhost:8000 \
        --endpoint-type chat \
        --streaming \
        --request-rate 100 \
        --arrival-pattern gamma \
        --arrival-smoothness $smoothness \
        --benchmark-duration 60 \
        --output-dir results/smoothness_$smoothness
done

Warmup with Stable Pattern, Profile with Realistic

Use constant arrivals during warmup, then realistic patterns for profiling:

aiperf profile \
    --model your-model \
    --url localhost:8000 \
    --endpoint-type chat \
    --streaming \
    --request-rate 100 \
    --arrival-pattern gamma \
    --arrival-smoothness 0.8 \
    --warmup-arrival-pattern constant \
    --warmup-duration 30 \
    --benchmark-duration 120

CLI Reference

Option Type Default Description
--arrival-pattern str poisson Pattern for request arrivals: constant, poisson, gamma
--arrival-smoothness float None Gamma smoothness: <1 = bursty, 1 = Poisson, >1 = smooth. Defaults to 1.0 when using gamma pattern.
--warmup-arrival-pattern str Inherits Override pattern for warmup phase

Constraints:

  • --arrival-pattern requires --request-rate to be set
  • --arrival-smoothness only applies when --arrival-pattern gamma
  • Cannot use with --user-centric-rate (deterministic per-user scheduling)
  • Cannot use with --fixed-schedule (timestamp-based scheduling)

Pattern Selection Guide

Goal Pattern Smoothness
Reproducible baseline constant N/A
Realistic traffic simulation poisson N/A
Match vLLM benchmark gamma Same as vLLM --burstiness
Stress test burst handling gamma 0.3 - 0.7
Reduce measurement noise gamma 2.0 - 5.0
Maximum throughput N/A (burst mode) N/A

Understanding the Math

For those who want to understand the statistical properties:

Pattern Distribution Mean Variance CV (Coeff. of Variation)
Constant Degenerate 1/λ 0 0
Poisson Exponential 1/λ 1/λ² 1
Gamma(k) Gamma 1/λ 1/(k·λ²) 1/√k

Where λ = request rate and k = smoothness.

  • CV (Coefficient of Variation) = standard deviation / mean
  • Lower CV = more regular arrivals
  • Gamma with k=1 equals Poisson (CV=1)
  • As k→∞, Gamma approaches Constant (CV→0)

Related Documentation