Skip to content

Tool Selection Guide ​

Category: Architecture

Use this guide to select the right tool for your Azure Local performance validation scenario. The decision flowchart covers the most common questions; the comparison table below it covers all dimensions.


Decision Flowchart ​

mermaid
flowchart TD
    A([What are you testing?]) --> B{Is it network throughput\nor latency?}
    B -->|Yes| C([iPerf3])
    B -->|No| D{Is it storage\nI/O performance?}
    D -->|Yes - block device\nbenchmark| E([fio])
    D -->|Yes - application\nI/O patterns| F{Is it a SQL\nworkload?}
    F -->|Yes| G([HammerDB])
    F -->|No| H([stress-ng io-stress])
    D -->|No| I{Is it CPU or\nmemory stress?}
    I -->|Yes| J([stress-ng cpu/memory])
    I -->|No| K{Is it full VM\nworkload simulation?}
    K -->|Yes| L([VMFleet])
    K -->|No| M([Consult team])

Tool Comparison Matrix ​

DimensionfioiPerf3HammerDBstress-ngVMFleet
Primary purposeBlock device I/O benchmarkingNetwork throughput & latencySQL database benchmarkingOS-level stress (CPU/memory/I/O)Full VM workload simulation
Target OSLinuxLinux / WindowsWindowsLinuxWindows (HCI host)
ProtocolPOSIX file I/OTCP / UDPTDS (SQL Server), libpq (PostgreSQL)POSIX / kernel syscallsHyper-V VM workload
Output formatJSONJSONHammerDB log + parsed JSONYAML + parsed JSONCSV / JSON
Profile count5323N/A (config-driven)
Install methodAnsible (Install-Fio.ps1)apt / dnfPowerShell remoting (Install-HammerDB.ps1)apt / dnfInstall-VMFleet.ps1
CI/CD readyYesYesYesYesYes
Monitoring alerts7 rules6 rules7 rules6 rules(shared PerfMon)
Key metricIOPS, throughput MB/s, latency P99Throughput MB/s, jitter msNOPM, TPMbogo-ops/secVM IOPS, CPU%
Parallelises across nodesYes (all nodes simultaneously)Yes (pairs / mesh)Yes (per-node DB instance)Yes (all nodes simultaneously)Yes (VM fleet distributes load)

Scenario Examples ​

"We want to know if our RDMA storage network is healthy after hardware replacement." ​

→ Use iPerf3 (mesh profile)

Mesh throughput between all node pairs will expose any link degraded from 10GbE to 1GbE, misconfigured MTU, or failed SFP.

"We need to validate that our NVMe SSDs meet the IOPS requirement for a new VM workload." ​

→ Use fio (random-read + random-write profiles)

fio directly benchmarks the block device from inside a VM, giving you raw IOPS and P99 latency that you can compare directly against the SSD datasheet and SLA requirement.

"We're deploying SQL Server on Azure Local and want to verify it can handle our TPC-C equivalent load." ​

→ Use HammerDB (tpc-c profile)

HammerDB simulates OLTP database operations and reports NOPM (New Orders Per Minute), the standard TPC-C throughput metric.

"We want to confirm that our AX nodes can sustain CPU load for 5 minutes without thermal throttling." ​

→ Use stress-ng (cpu-stress profile)

Saturates all logical CPUs and reports bogo-ops/sec; the stressng_cpu_throttling alert fires if frequency drops below 80%.

"We want to understand how many VMs our cluster can host before storage IOPS degrades." ​

→ Use VMFleet

VMFleet deploys a configurable fleet of VMs with CDB workload simulation and measures cluster-wide storage throughput as VM density increases.


Combining Tools ​

For production validation, run tools in sequence:

  1. iPerf3 mesh — Confirm network fabric health before any other testing
  2. stress-ng cpu-stress — Verify thermal and BIOS power profile settings
  3. fio sequential + random — Baseline storage performance per node
  4. HammerDB tpc-c — Validate SQL workload under realistic OLTP load
  5. VMFleet — Full-system capacity proving

Each tool writes results to its own logs\<tool>\<RunId>\ directory; reports can be generated independently after each phase completes.

Released under the MIT License.