In IBM PowerVM, the hypervisor tracks CPU utilization per partition (LPAR) using hardware-assisted accounting + fine-grained scheduling counters. The goal is to know exactly how much physical CPU time each LPAR actually consumes vs. what it is entitled to.

🚀 Core Idea

Every time a virtual processor from an LPAR runs on a physical core, the hypervisor:

👉 Starts a timer → runs the workload → stops the timer

That elapsed time is accounted to that specific LPAR

⚙️ How CPU Utilization Is Tracked

1. Dispatch-Time Accounting (Core Mechanism)

• When a virtual processor is dispatched:
  • Hypervisor records start timestamp
• When it is preempted:
  • Records end timestamp

👉 CPU usage = accumulated run time across all dispatches

2. Per-Partition Counters

The hypervisor maintains counters such as:

• Entitled CPU time (guaranteed share used)
• Consumed CPU time (actual usage)
• Idle time (when LPAR had no work)
• Wait time (ready but not scheduled)

👉 These are tracked continuously

3. Physical Processor Mapping

• Tracks which physical cores and SMT threads:
  • Each LPAR runs on

👉 Enables:

• Accurate per-core utilization
• Load balancing decisions

4. Timebase / Hardware Registers

POWER processors provide:

• High-resolution timebase registers

👉 Hypervisor uses these to:

• Measure CPU time precisely (nanosecond-level granularity)

5. Virtual Processor State Tracking

Each vCPU can be in states like:

• Running
• Ready (waiting)
• Idle

👉 Hypervisor tracks transitions between these states to compute:

• CPU ready time
• Utilization vs demand

6. Shared Processor Pool Accounting

• Usage is also tracked at:
  • Pool level
  • Partition level

👉 Helps enforce:

• Fair sharing
• Entitlement guarantees

📊 Key Metrics Derived

🔹 Entitled Utilization

• % of guaranteed CPU actually used

🔹 Physical CPU Consumption

• Actual CPU cycles consumed

🔹 Virtual Processor Utilization

• How busy each vCPU is

🔹 CPU Ready Time

• Time spent waiting for CPU

👉 Critical for detecting:

• Overcommit issues

⚡ Example

🔗 Interaction with System Components

Scheduling

• Data feeds into:
  • Dynamic CPU allocation
  • Load balancing decisions

VIOS Impact

• Accurate tracking ensures:
  • Virtual I/O Server (VIOS) gets enough CPU
• Prevents I/O bottlenecks

Performance Tools

Metrics are exposed to tools like:

• topas
• nmon
• lparstat

👉 These tools read hypervisor-provided counters

🧠 Key Insight

CPU utilization tracking in PowerVM is:

time-based, hardware-assisted, and continuously updated per partition

—not just an estimate.

🎯 Why It Matters

• Enables fair CPU sharing
• Helps detect bottlenecks
• Supports capacity planning
• Ensures SLA compliance

🔑 Final Takeaway

PowerVM tracks CPU utilization per partition by:

• Measuring actual execution time on physical CPUs
• Using high-resolution hardware timers
• Maintaining per-LPAR accounting counters

👉 This precise tracking is what makes dynamic scheduling, overcommit handling, and fairness possible.