Performance methodology
peryx uses the same cache path across ecosystem owners. A miss streams bytes from the upstream to the client and content store. Concurrent requests for one missing object share that fetch. A warm read serves content-addressed bytes from local storage.
Cold path
The server forwards chunks as they arrive and writes the same chunks to storage. It verifies the digest before making new metadata visible. The client pays upstream transfer time, one proxy hop, and stream processing; it does not wait for a second full copy into storage.
Warm path
Content addressing stores one copy of equal bytes referenced by multiple entities. Single-flight fetches prevent a client burst from multiplying an upstream transfer. Local placement removes wide-area round trips from dependent metadata and artifact requests.
Benchmark controls
Each workload runs in independent rounds. A round starts the server with empty state, measures the cold operation, then measures the warm operation against the populated cache. Process groups and client caches reset between rounds.
Tables report the median and coefficient of variation. Network-bound rows provide context but do not gate regressions. Local workloads compare revisions on the same host and power state.
Request-load tests use an open-loop schedule. Each request keeps its intended send time, so a stall contributes its full delay to tail latency. A coordinated-omission-safe histogram records percentiles.
The harness measures wall time, throughput, latency, process-tree CPU, and peak resident memory. It records storage, memory, and loopback baselines from the same machine. Those baselines identify client subprocess or filesystem limits that might otherwise look like server behavior.
Host controls
Each benchmark report records its host and storage layout beside its result tables. Each baseline discards a warm-up sample and reports the median and spread of five samples. The harness places server state, client caches, and scratch files on the system temporary volume, separate from the repository checkout.
Memory, disk, and minimal HTTP rows provide scales for serving results. The minimal server measures socket and system-call cost. Disk results identify storage limits, but warm artifact reads can come from the operating system page cache.
Reading results
Cold results include upstream and network variance. Warm results isolate local serving and client work. A complete client operation can hide server differences behind resolution, decompression, verification, or installation, so each ecosystem benchmark also measures protocol endpoints and large-artifact transfer.
Throughput above the deployment network rate compares host efficiency rather than user-visible speed. Normalize CPU by completed work. Failed operations do not count as throughput.