goodput

Measurement harness for the sliding window attention premium in the vLLM TPU Ragged Paged Attention v3 kernel: per layer decode cost, block size control, throughput, and goodput for Gemma 4 31B on TPU v6e.

One model, many configs, real measured TTFT/TPOT/goodput under load, with a written tradeoff analysis and a config recommendation per SLO.

One model, many configs, real measured TTFT/TPOT/goodput under load, with a written tradeoff analysis and a config recommendation per SLO.

Evidence-first AI serving capacity tests for HTTP/SSE/WebSocket, measured as SLO-qualified goodput.

Simulates LLM serving admission control strategies (None, MaxConcurrency, TokenBudget, SLOAware, Predictive) across 135 configurations. Key findings: tight token budget (4096) achieves 10x better goodput than loose budget; predictive control is the only strategy maintaining SLO compliance at 100ms; SLOAware is fundamentally unstable under sustained

Measuring the KV-transfer tax of disaggregated prefill/decode LLM serving (vLLM + NIXL, 4x A10G): on PCIe-only hardware, disagg loses to plain data-parallel replication — measured, committed, reproducible.