# Cloud Run performance and cold starts

> How fast Cloud Run starts and serves: cold starts measured, minimum instances, startup CPU boost, concurrency, connection pools, GPUs, Rust.

<https://exodus.elim.io/blog/topic/performance>

These guides measure how fast a Cloud Run app starts and answers, and what each speed costs. They cover cold starts as measured across platforms and in our own starts, minimum instances and startup CPU boost, how many requests one instance takes, connection pools to Cloud SQL, GPUs for your own model, and what a Rust service does in its first second.

- [Connection pooling on Cloud Run: sizing the pool, Cloud SQL's 25 connections, PgBouncer or Managed Connection Pooling](https://exodus.elim.io/blog/connection-pooling-on-cloud-run.md): Each Cloud Run instance opens its own pool, and the smallest Cloud SQL machine takes 25 connections. How to size pools per framework, and when to use Managed Connection Pooling or PgBouncer.
- [GPUs on Cloud Run: running your own model, what it costs, and where the start goes](https://exodus.elim.io/blog/gpus-on-cloud-run.md): Cloud Run attaches an NVIDIA L4 or RTX PRO 6000 to an instance and scales it to zero. What a deploy needs, a sample model image we built and timed, where the cold start goes, and the cost.
- [Minimum instances and startup CPU boost: what a cold start costs, and what it costs to skip it](https://exodus.elim.io/blog/min-instances-and-startup-cpu-boost.md): Startup CPU boost makes every start faster for a fraction of a cent; a minimum instance skips the wait for about $8 to $10 a month. How each works, how to measure first, the costs and the traps.
- [Rust on Cloud Run: the container, the build, the first second, and the bill](https://exodus.elim.io/blog/rust-on-cloud-run.md): How we run Rust on Cloud Run: $PORT and SIGTERM, a two-stage Dockerfile for an 11 MB image, where to compile, why 256 MiB means the first generation, start times, and three setups' bills.
- [How fast Cloud Run wakes up: cold starts, measured](https://exodus.elim.io/blog/cloud-run-cold-start-measured.md): An idle Cloud Run app stops, and the next visitor waits for a start. What a study across eight platforms measured, what 703 of our own starts took, and how to measure and shorten yours.
- [Concurrency on Cloud Run: how many requests one instance takes](https://exodus.elim.io/blog/concurrency-on-cloud-run.md): One setting decides how many requests share a Cloud Run instance. What it limits, how the autoscaler reads it, why your web server decides the real number, and its effect on the bill and database.