Synthesizing Golden Traces for Realistic KV Cache Evaluation in Agentic Inference Workloads
DDN, Wednesday, July 29th, 2026
DDN shows why synthetic benchmarks misrepresent KV cache behaviour and how to build realistic traces.
Sizing inference clusters and storage tiers is difficult largely because conventional synthetic benchmarking tools produce access patterns that do not resemble real workloads.
DDN focuses on KV cache behaviour in agentic inference, where multi-turn conversations and tool calls create reuse patterns that random synthetic traces miss entirely.
The post describes synthesizing golden traces that reproduce realistic cache hit distributions. Getting this right changes capacity conclusions materially, because cache-friendly workloads need very different storage than the synthetic baseline suggests.
The work targets teams sizing infrastructure for agentic inference.