VMware ESXi • Kernel Internals

Reading an ESXi Host Configuration: Kernel Parameters, NUMA, & Power Profiles in Production

When onboarding or performing a health check on an ESXi cluster, the default vSphere client dashboard only scratches the surface. Critical cluster performance bottlenecks—such as cross-socket memory access, CPU frequency scaling throttling, and interrupt pinning—stem from ESXi host-level kernel defaults.

In this guide, we will step through inspecting and tuning an ESXi host from the command line using esxcli and vsish, examining the exact metrics that determine whether your virtual machines experience deterministic low latency.

1. Inspecting NUMA Topology & Memory Locality

Modern multi-socket servers (Intel Xeon Scalable and AMD EPYC) partition CPU cores and memory channels into Non-Uniform Memory Access (NUMA) nodes. If a VM's vCPU executes on Socket 0 but its memory resides on Socket 1, every memory read incurs high inter-socket UPI/Infinity Fabric latency.

ESXi Shell — NUMA Inspection Bash
# List physical NUMA nodes and memory distribution esxcli hardware cpu global get vsish -e get /hardware/numa/numNodes vsish -e get /hardware/numa/nodeList
Key Production Insight Keep VM vCPU counts under or matching the physical NUMA node size whenever possible. When a VM spans multiple NUMA nodes (vNUMA), verify in vmid.log that ESXi correctly initialized matching virtual NUMA topologies.

2. Power Management & DPM Governor Penalties

By default, many server BIOS vendors and ESXi default to "Balanced" power profiles. While this conserves wattage under low load, entering and exiting CPU C-states (C1E, C6) incurs measurable microsecond sleep exit latency:

Query and Set High Performance Governor Bash
# Check current power policy esxcli system settings advanced list -o /Power/CpuPolicy # Set to static High Performance (avoids CPU clock frequency stepping) esxcli system settings advanced set -o /Power/CpuPolicy -s "High Performance"

3. Critical Kernel Parameters Under Production Load

For clusters hosting high-throughput databases or low-latency microservices, three kernel parameters warrant explicit review:

  • Disk.SchedNumReqOutstanding (DSNRO): Controls the queue depth limit per VM for a shared datastore. Default is 32; on all-flash SANs, increase this to 64 or 128.
  • Net.IntrTxThreshold: Adjusts network transmit interrupt coalescing on high-speed 25/100GbE vmnics.
  • Mem.AllocGuestLargePage: Controls whether the hypervisor attempts 2MB huge page allocation for guest memory mappings.

4. Automated Host Audit Script

You can run this simple one-liner during host commissioning to capture a snapshot of vital host settings:

PowerCLI Host Baseline Query PowerShell / PowerCLI
Get-VMHost | Select Name, Version, Build, @{N="PowerPolicy"; E={(Get-VMHostAdvancedConfiguration -VMHost $_ -Name "Power.CpuPolicy").Values}}, @{N="CoresPerSocket"; E={$_.ExtensionData.Hardware.CpuInfo.NumCpuCores / $_.ExtensionData.Hardware.CpuInfo.NumCpuPackages}} | Format-Table -AutoSize

5. Summary & Next Steps

Reviewing an ESXi host before adding it to production DRS clusters ensures that resource calculations reflect consistent hardware availability across all nodes. In our next teardown, we will explore packet scheduling through the vSphere Distributed Switch (vDS).

Moinul Haque

Written by Moinul Haque

Practicing infrastructure architect documenting VMware systems engineering and production hypervisors.