Skip to content

VMFleet Workload Profiles ​

Tool: VMFleetCategory: Tool Guide

This guide covers the predefined workload profiles for VMFleet testing and how to execute them.

Built-in Profiles ​

ProfileDescription
GeneralBalanced 70/30 read/write, 70% random, 8K blocks — general purpose validation
Peak IOPS100% random 4K reads — maximum IOPS measurement
VDIMixed small block I/O simulating Virtual Desktop Infrastructure patterns
SQL OLTP8K random read/write mix simulating SQL Server OLTP workloads
Sequential Throughput100% sequential 512K writes for maximum MB/s measurement
Day2SmokeThe General workload shortened to about 15 minutes for a readiness gate: see Day-2 smoke
FailureInjectionThe General workload held for 20 minutes so one injected failure can be measured. Opt-in: see failure injection

Running Tests ​

Single Profile ​

powershell
.\tools\vmfleet\scripts\Start-VMFleetTest.ps1 `
    -ProfilePath "tools/vmfleet/config/profiles/general.yml" `
    -DurationSeconds 300

Profile Sweep ​

powershell
.\tools\vmfleet\Invoke-VMFleetPipeline.ps1 `
    -ClusterConfigPath "config/clusters/my-cluster.yml" `
    -ProfilePaths @(
        "tools/vmfleet/config/profiles/general.yml",
        "tools/vmfleet/config/profiles/peak-iops.yml",
        "tools/vmfleet/config/profiles/sql-oltp.yml",
        "tools/vmfleet/config/profiles/vdi.yml",
        "tools/vmfleet/config/profiles/sequential-throughput.yml"
    )

Profile Configuration ​

Profiles are defined in tools/vmfleet/config/profiles/:

yaml
# tools/vmfleet/config/profiles/general.yml
profile:
  name: "General"
  description: "Balanced read/write workload for general validation"
  parameters:
    block_size: "8k"
    write_ratio: 30
    random_ratio: 70
    outstanding_io: 8
    threads_per_vm: 2
    duration_seconds: 300
    warmup_seconds: 60

Profile Parameters ​

ParameterDescriptionExample
block_sizeI/O block size4k, 8k, 64k
write_ratioPercentage of I/O that is writes (0–100)30 for 70/30 read/write
random_ratioPercentage of I/O that is random (0–100)70 for mostly random
outstanding_ioI/O queue depth per thread8
threads_per_vmDiskSpd threads per fleet VM2
duration_secondsTest duration (excluding warmup)300
warmup_secondsWarmup before measurement starts60

DiskSpd Parameter Mapping ​

Profile keyDiskSpd switchMeaning
block_size-bI/O block size
write_ratio-wWrite ratio percentage
random_ratio-rRandom I/O ratio percentage
outstanding_io-oQueue depth per thread
threads_per_vm-tThreads per VM
duration_seconds-dMeasurement duration
warmup_seconds-WWarmup duration

Expected Thresholds ​

ProfilePrimary metric focusTypical pass criteria
GeneralBalanced IOPS + latencyStable mixed IOPS and latency under target
Peak IOPSMax read IOPSHighest sustained random read IOPS
SQL OLTPRead/write DB patternLow latency under mixed random load
VDILight mixed desktop loadPredictable latency across many VMs
Sequential ThroughputMB/sHighest sustained throughput

Profile Selection Guide ​

  • Use peak-iops.yml for storage ceiling discovery.
  • Use sql-oltp.yml for database-aligned validation.
  • Use vdi.yml for desktop density readiness checks.
  • Use sequential-throughput.yml for bulk transfer throughput checks.
  • Use general.yml for a balanced baseline run.

Collecting Results ​

After tests complete, parse DiskSpd XML output and aggregate metrics:

powershell
.\src\solutions\vmfleet\scripts\Collect-VMFleetResults.ps1 `
    -ResultsPath "results/current-run/"

Next Steps ​

Read this before you compare results ​

The CSV in-memory read cache can make these tools look worse ​

Learn says: "Certain microbenchmarking tools like DISKSPD and VM Fleet may produce worse results with the CSV in-memory read cache enabled than without it." VMFleet by default creates one 10 GiB VHDX per VM and then issues uniformly random reads and writes against them. Unlike a real workload, the reads follow no predictable or repeated pattern, so the in-memory read cache is not effective and "just incurs overhead" (Use the CSV in-memory read cache).

What to do with that:

  • Record the CSV cache size next to every result ((Get-Cluster).BlockCacheSize, in MiB per server; the Azure Local default is 1 GiB).
  • Compare runs only when the cache setting is the same. A result taken with the cache off is not comparable with one taken with it on.
  • Do not conclude that the cache is bad for your workloads from a synthetic uniform-random test. Learn says it is most effective for read-intensive workloads such as VDI, and that writes cannot be cached in memory.
  • Do not change the cache setting on a production cluster to chase a benchmark number.

Validate with DiskSpd from representative VMs, including under failure ​

The Azure Well-Architected Framework guidance for Azure Local recommends using DiskSpd, application benchmarks or SAN vendor tools "from representative VMs to validate latency, throughput, and IOPS under normal and intended failure conditions" (Architecture best practices for Azure Local). The Azure Local baseline reference architecture says synthetic tools such as DiskSpd or VMFleet "isolate storage subsystem constraints and establish a repeatable performance baseline", and that the application itself should be load-tested to confirm end-to-end targets (baseline architecture). See also DiskSpd overview.

This repository supports the failure part with an opt-in helper: failure injection.

Released under the MIT License.