Problems / LLM serving, medium
Serving is split into prefill and decode pools (disaggregated serving). Prompts got longer: the prefill pool can’t keep up, and time to first token suffers.
Free to open. Runs in your browser, no sign-up.
Success ≥ 99%, p99 ≤ 5 s.
300 requests a second for 30 seconds: 100% request (reads).
What each component does in this design, and what calls it. The simulator mimics each one from these settings.
| Component | Its job here | Settings |
|---|---|---|
| Load balancerLoad balancer | Traffic enters here.Calls AI app.Layer 4 balancers route by IP and port; layer 7 balancers read HTTP and can route by path or header. | – |
| AI appApp service | Load balancer forwards to it.Calls Prefill GPU pool, Decode GPU pool.Little's law: throughput ≈ concurrency ÷ latency. 200 slots at 50 ms serve about 4,000 req/s. | 3 replicas × 500 slots · timeout 30 s |
| Prefill GPU poolPrefill GPU pool | AI app calls it.Splitting prefill (reading the prompt) from decode (writing the answer) onto separate GPUs stops long prompts from stalling everyone else’s output. | 32 slots |
| Decode GPU poolDecode GPU pool | AI app calls it.Decoding is memory-bound: each request holds its KV cache in GPU memory for as long as it generates. | 1200 slots |
Try it first. Open one hint at a time if you get stuck.
Which pool is red?
Prefill work grows with prompt length; decode with output length.
Add capacity to the prefill pool.