Cloud Infrastructure · Cluster
Compute Models
Instances, functions, and scheduling units.
- 01 The Anatomy of a Virtual Instance: vCPU Threads, Guest Memory, and Device Models What an instance is composed of beneath the API: a set of host threads, a second-level page table, and a set of emulated or shared devices.
- 02 Steal Time and the Mechanics of CPU Oversubscription How a hypervisor shares physical cores among more vCPUs than it has, how the guest learns it was descheduled, and what the accounting can and cannot reveal.
- 03 Cold Start Anatomy: What Happens Between a Request and the First Instruction The stages a function invocation passes through before user code runs, which stages dominate the delay, and why warm pools change the shape of the curve.
- 04 Concurrency Per Environment and the Unit a Function Platform Scales Why serverless platforms scale environments rather than requests, how per-environment concurrency changes the arithmetic, and where the queue lives when the limit is reached.
- 05 Requests, Limits, and the Two Different Jobs They Do A request is a scheduling promise and a limit is an enforcement ceiling; conflating them produces nodes that look full but throttle, and pods that are evicted for reasons nobody expected.
- 06 Interruptible Capacity: Eviction Notices, Checkpointing, and the Cost of Being Preempted What a preemptible instance actually promises, how the notice arrives, and which workload properties determine whether the discount is worth the interruption.
- 07 Autoscaling Control Loops: Why the Metric Lags the Demand The feedback loop behind replica count: measurement windows, stabilization, and the delay stack that makes a fast scaler overshoot and an accurate one react too late.
- 08 Placement as Bin Packing: Multi-Dimensional Fit and the Fragmentation It Creates Why a scheduler with free cores and free memory can still refuse a pod, how multi-dimensional packing produces unusable holes, and which scoring functions trade density against headroom.
- 09 Gang Scheduling: All-or-Nothing Allocation for Distributed Workloads Why a training job or MPI rank set cannot be placed one member at a time, how reservation schemes avoid deadlock, and what each scheme costs the rest of the cluster.
- 10 What CPU Quotas Cannot Bound: Cache, Bandwidth, and SMT Contention Between Tenants Fair-share scheduling accounts for time, not for shared microarchitectural state; the resources that leak between co-tenants and the placement choices that limit the leak.