Beyond Your Bill — selected guide
Monitoring for efficient cluster binpacking in GKE
Eleven episodes on resource ownership, billing analysis, budgets, commitments, and GKE efficiency
Eleven 2019–2021 episodes explain who owns cloud spending, how to analyze it, and where commitments and GKE sizing can reduce waste. Budget alerts notify teams; they do not impose a spending cap.
Google Cloud Tech 10 of 11
In this collection Browse 11 summaries 10 of 11
Bin packing means fitting workloads onto cluster machines so less capacity sits unused. This 2021 Google Kubernetes Engine (GKE) walkthrough explains why CPU, memory, and pod counts must be examined together: a machine can be full by one measure and wasteful by another.
Key Points Covered
- Inspect node pools, not only cluster averages. A node pool is a group of machines assigned to workloads. If one pool is poorly sized for its applications, its unused capacity may be hidden in the cluster-wide average. 01:06–02:09(opens in a new tab)
- Compare four measurements. Allocatable capacity is what remains available to workloads after system components take their share. Requests are what pods ask for, recommendations are the platform's suggested requests, and usage is what pods actually consume. The gaps show where to investigate sizing. 02:09–03:11(opens in a new tab)
- Check memory before cutting a CPU-heavy machine. In the example, workloads use little CPU but most available memory. The presenter suggests investigating machines with the same memory and fewer CPUs, rather than simply shrinking every resource. 03:11–04:14(opens in a new tab)
- Pod limits can leave CPU and memory unused. Limiting pods per node can conserve IP addresses because each pod needs one. But reaching that pod limit prevents further packing even if the machine still has spare compute resources. 03:11–05:27(opens in a new tab)
Full video: https://www.youtube.com/watch?v=lPSdlTIJ00c(opens in a new tab)