You already know: prefill is compute-bound and decode memory-bound (Inference fundamentals); chunked prefill reduces but doesn't remove their interference (Batching & scheduling); KV cache size per token (KV cache engineering); goodput under SLOs (Measurement & capacity); network bandwidth and latency (D1 Networking).
Why this matters for Atlas. On a shared GPU, prefill and decode fight: a prefill step delays every decoding sequence (TPOT spikes), and decode-heavy batches delay new prompts (TTFT grows). Every scheduler knob trades one against the other. Disaggregated serving stops the fight by running them on separate GPU pools: prefill workers process prompts and ship the resulting KV cache over a fast network to decode workers, which generate the tokens. Each pool can be sized, batched and even built from different hardware for its own phase. This is how several of the largest LLM services run today. This lesson explains when the split pays, what the KV transfer costs, and how to size the two pools.
1. The interference problem, measured
The question. Chunked prefill helped. Why isn't it enough?
Start from nothing. With colocated serving, each step mixes some prefill tokens with the decode batch. Every prefill token in a step makes that step longer for all decoding sequences, so TPOT rises with prefill load. You can cap prefill per step (small chunks) to protect TPOT, but then long prompts take more steps and TTFT rises. One budget, two SLOs: you can't fully satisfy both when prompts are long and traffic is high.