Hypervisor CPU scheduling is all about deciding which virtual machine (LPAR/VM) runs on a physical core at any given moment—and for how long—while adapting continuously to workload demand.

In platforms like IBM PowerVM, this is done with dynamic, policy-driven scheduling rather than fixed CPU assignment.

🚀 Core Idea

Instead of giving each VM a fixed CPU, the hypervisor:

• Tracks demand (load)
• Applies entitlements and priorities
• Dynamically allocates CPU cycles in small time slices

👉 Result: efficient utilization + fairness + performance isolation

⚙️ Key Scheduling Concepts

1. Entitlement (Guaranteed CPU)

• Each LPAR is assigned a minimum CPU capacity
  • Example: 0.5 CPU, 2 CPUs, etc.

👉 The hypervisor ensures:

• This capacity is always available (if needed)

2. Shared Processor Pool

• CPUs are placed in a shared pool
• Multiple LPARs draw from the same pool

👉 Enables:

• Dynamic redistribution of unused CPU

3. Capped vs Uncapped Mode

Capped

• LPAR cannot exceed its entitlement
• Predictable but less flexible

Uncapped

• LPAR can borrow unused CPU cycles from others

👉 Key for dynamic allocation

🔄 How Dynamic Allocation Works

Step-by-Step Scheduling Flow

1. Monitor Demand
   • Hypervisor checks:
     • Run queues
     • CPU utilization
     • Thread readiness
2. Assign Time Slices
   • CPU time divided into very small slices (milliseconds)
3. Dispatch Threads
   • LPAR threads mapped to physical cores (SMT threads)
4. Rebalance Continuously
   • If one LPAR is idle:
     • Its cycles are redistributed
   • If another LPAR spikes:
     • It gets extra CPU (if uncapped)

⚡ Advanced Mechanisms

🔹 Weight-Based Scheduling

• Each uncapped LPAR has a weight value
• Higher weight → more share of extra CPU

👉 Example:

• LPAR A (weight 200) gets more spare CPU than LPAR B (weight 100)

🔹 Simultaneous Multithreading (SMT) Awareness

• POWER CPUs run multiple threads per core
• Scheduler decides:
  • How many threads per LPAR
  • Whether to spread or pack workloads

🔹 Dispatch Groups

• Threads grouped for efficient scheduling
• Reduces context-switch overhead

🔹 Affinity & Cache Locality

• Scheduler tries to:
  • Keep workloads on same core/cache
• Improves performance by:
  • Reducing cache misses

📊 Example Scenario

👉 Outcome:

• C gets exactly 1 CPU
• B uses less → frees cycles
• A borrows extra → may get 3–4 CPUs dynamically

🧠 Why This Works Well

✅ Benefits

• High CPU utilization (minimal idle time)
• Adapts instantly to workload spikes
• Supports consolidation of many workloads
• Ensures fairness + priority control

⚠️ Trade-offs

• Possible CPU contention if oversubscribed
• Latency-sensitive apps may see jitter
• Requires careful tuning:
  • Entitlements
  • Weights
  • Pool sizes

🔗 Interaction with I/O & Virtualization

• Works alongside:
  • Virtual I/O Server (VIOS) (needs guaranteed CPU for I/O)
• Poor CPU scheduling can:
  • Impact I/O latency
  • Create bottlenecks

🎯 Key Insight

Hypervisor scheduling dynamically allocates CPU cycles by combining:

• Guaranteed minimums (entitlement)
• Opportunistic sharing (uncapped mode)
• Priority weighting

👉 Turning a fixed set of CPUs into a flexible, demand-driven compute pool