DocumentationGradual ramping lets you increase concurrency and request rate smoothly over time, rather than jumping to full load immediately. This prevents overwhelming your server at benchmark start.
When a benchmark starts, immediately hitting your target load can cause problems:
Problems without ramping:
Benefits of ramping:
AIPerf supports ramping three dimensions:
Each ramps from a low starting value up to your target over the specified duration.
Gradually increase from 1 concurrent request to 100 over 30 seconds:
Sample Output (Successful Run):
What happens:
Gradually increase from a low starting rate to 100 QPS over 60 seconds:
Sample Output (Successful Run):
What happens:
The starting rate is calculated proportionally: start = target * (update_interval / duration). With default settings (0.1s updates), ramping to 100 QPS over 60 seconds starts at ~0.17 QPS (not zero).
Ramp both dimensions together for maximum control:
Sample Output (Successful Run):
Both ramp in parallel, reaching their targets at 30 seconds.
For long-context workloads, you may want to limit how many requests are in the “prefill” phase (processing input tokens) simultaneously. This prevents memory spikes from multiple large prompts being processed at once.
Sample Output (Successful Run):
This limits prefill to 20 concurrent requests (ramped over 20 seconds), while allowing up to 100 total concurrent requests.
Each phase can have its own ramp settings. Warmup uses --warmup-* prefixed options:
Ramp slowly to avoid overwhelming the server, then sustain full load:
The 60-second ramp gives the server time to allocate resources (~8 new connections per second).
Limit memory spikes from large prompts with prefill concurrency:
Only 5 requests process their 32K input tokens simultaneously, preventing KV-cache OOM.
Find your server’s limits by ramping slowly and watching for degradation:
Watch latency and throughput metrics as load increases. When latency spikes or errors appear, you’ve found the limit.
Concurrency increases by +1 at evenly spaced intervals:
Each step allows one more concurrent request immediately.
Rate updates continuously (every 0.1 seconds by default):
This creates smooth traffic curves without sudden jumps.
Key behaviors:
