diskspd

DiskSpd

DiskSpd is a storage performance tool from the Windows, Windows Server and Cloud Server Infrastructure engineering teams at Microsoft. Please visit https://github.com/Microsoft/diskspd/wiki for updated documentation.

In addition to the tool itself, this repository hosts measurement frameworks which utilize DiskSpd. The initial example is VM Fleet, used for Windows Server Hyper-Converged environments with Storage Spaces Direct.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.

Releases

The Releases page includes pre-compiled binaries (ZIP) and source code for the most current releases of the DiskSpd tool. The latest update to DiskSpd can always be downloaded from https://github.com/Microsoft/diskspd/releases/latest/download/DiskSpd.zip (aka https://aka.ms/getdiskspd).

What's New?

DISKSPD

DISKSPD 2.3 9/3/2026

NOTE: two default changes may require rebaselining. Thread affinity now orders P-cores before E-cores within each group by default (-aup restores unordered assignment), which changes placement on heterogeneous systems. I/O buffers are now separated by PDE cache line by default (-bsp; use -bsn for the prior system-default allocation).

• New IoRing I/O path (-u[n]<batchsize>[p]) with registered buffers by default and configurable submission batching.
• New BypassIO modes: -Sy allows partial bypass and -SY requires full bypass.
• New topology-aware affinity controls:
  • -ac assigns one thread per physical core before using SMT siblings.
  • -as spans processor groups before filling each group.
  • -ap, -aP, -ae, and -aE control P-core/E-core ordering; -aup disables efficiency ordering.
  • Explicit affinity supports CPU ranges and whole-group selection.
• New processor name and cache topology reporting in text and XML results.
• New controls for I/O buffer separation (-bs) and completion-port dequeue depth (-oc).
• New system-information-only output (-Rs).
• XML profiles and results support the new I/O, affinity, buffer, and completion settings.
• Visual Studio projects now share common settings, can build Win32, x64, and ARM64, and integrate TAEF tests with Test Explorer.

The binary release contains amd64 and arm64 builds. The minimum supported operating system is Windows 10, version 1803 (RS4). IoRing and BypassIO require Windows 11 or later; BypassIO also requires NTFS.

DISKSPD 2.2 6/3/2024

NOTE: changes to the asynchronous IO loop will require rebaselining results with queue depths greater than 1. The new design drains the completion queue more aggressively, shrinking delays that impacted latency measurement especially on faster storage targeting higher queue depths. Latency measurement is improved at a very small cost to rates, as well as improving rates when latency measurement is not used (-D or -L) due to batch dequeue.

Smaller IO sizes will see the most change.

• New: Socket, NUMA, Core and Power Efficiency Class (big/little cores) added to processor topology reporting (XML and text output)
  • topology elements only displayed in text results when > 1 are present (e.g. multi-socket systems)
  • CPU numbering remains group relative, as is the new Core numbering
  • highest Power Efficiency Class is marked with a P suffix (this will usually be 1P v. 0)
  • NOTE: efficiency classes can have major impact; work on heterogenous systems must be aware of core properties in combination with thread affinity rules (see -a and -n )
• New: active power scheme reporting
• New: base target offset switch now allows range specification: -Bbase[:length]; replaces -B and -f
• post-run latency histogram processing now significantly faster
• verbose output is more consistent; includes actual warmup, measured and cooldown intervals v. expected

Binary release supports down to Windows 8 and Windows Server 2012; now uses the Universal CRT.

DISKSPD 2.1 7/1/2021

• New -g<n>i form allowing throughput limit specification in units of IOPS (per specified blocksize)
• New -rs<pct> to specify mixed random/sequential operation (pct random); geometric distribution of run lengths
• New -rd<distribution> to specify non-uniform IO distributions across target
  • pct by target percentage
  • abs by absolute offset
• New -Rp<text|xml> to show specified parameter set in indicated profile output form; works with -X XML profiles and conventional command line
• XML results/profiles are now indented for ease of review
• Text result output updates
  • now shows values in size units (K/M/G, and now TiB) to two decimals
  • thread stride no longer shown unless specified
  • -F/-O threadpool parameters shown
• XML profiles can now be built more generically
  • XML profiles can be stated in terms of templated target names (*1, *2), replaced in order from command line invocation
  • the command line now allows options alongside -X: -v, -z, -R and -W/-d/-C along with template target specs

DISKSPD 2.0.21a 9/21/2018

• Added support for memory mapped I/O:
  • New -Sm option to enable memory mapped I/O
  • New -N<vni> option to specify flush options for memory mapped I/O
• Added support for providing Event Tracing for Windows (ETW) events
• Included a Windows Performance Recorder (WPR) profile to enable ETW tracing
• Added system information to the ResultParser output

DISKSPD 2.0.20a 2/28/2018

• Changes that may require rebaselining of results:
  • New random number generator that may show an observable decreased cost
  • Switched to 512-byte aligned buffers with the -Z option to increase performance
• New -O option for specifying the number of outstanding IO requests per thread
• New -Zr option for per-IO randomization of write buffer content
• XML: Adds a new <ThreadTarget> element to support target weighting schemes
• Enhanced statistics captured from IOPS data
• Added support for validating XML profiles using an in-built XSD
• Added support for handling RAW volumes
• Updated CPU statistics to work on > 64-core systems
• Updated calculation and accuracy of CPU statistics
• Re-enable support for ETW statistics

DISKSPD 2.0.18a 5/31/2016

• update /? example to use -Sh v. deprecated -h
• fix operation on volumes on GPT partitioned media (:)
• fix IO priority hint to proper stack alignment (if not 8 byte, will fail)
• use iB notation to clarify that text result output is in 2^n units (KiB/MiB/GiB)

DISKSPD 2.0.17a 5/01/2016

• -S is expanded to control write-through independent of OS/software cache. Among other things, this allows buffered write-through to be specified (-Sbw).
• XML: adds a new <WriteThrough> element to specify write-through
• XML: <DisableAllCache> is no longer emitted (still parsed, though), in favor or <WriteThrough> and <DisableOSCache>
• Text output: OS/software cache and write-through state are now documented separately (adjacent lines)
• Latency histogram now reports to 9-nines (one part in one billion) in both text and XML output
• Error message added for failure to open write-content source file (-Z<size>,<file>)

VM Fleet

VM Fleet is a performance characterization and analyst framework for exploring the storage capabilities of Windows Server Hyper-Converged environments with Storage Spaces Direct.

VM Fleet 2.1.0.0 4/3/2024

• Support for Arc VM management (only applicable to clusters managed by Arc)
• Set-FleetRunProfileScript - produce a free-run script based on one of the defined workload profiles
• Watch-FleetCPU - new support for monitoring guest VCPU utilization (-Guest); can handle data outages
• Fix: performance counter handling now manages intermittent data dropouts (per conventional relog.exe)
• Fix: mid-run vm health check now handles the possibility of many vms taking longer than intended runtime to validate; early exit to avoid false failures
• Fix: ignore reboot required indication from cache layer when changing cache behavior; avoid false failure

VM Fleet 2.0.2.2 12/1/2021

• Fix cluster remoting issue during New-Fleet caused by 2.0.2.1 work
• Use timestamped logging in New-Fleet, simplify and de-colorize default output

VM Fleet 2.0.2.1 11/9/2021

• Fix cluster remoting issues in Move-Fleet and Get-FleetDataDiskEstimate
• Fix timing issue with Start-FleetSweep; always start from fleet pause to avoid triggering free run
• Use uniqueness to guarantee Start-FleetSweep runs profile in case of repeat

VM Fleet 2.0.2 11/2/2021

• Windows Server 2019/RS5 host operation now confirmed & supported
• Read cache warmup for HDD capacity systems should now be faster

Set-FleetPause will wait for VM responses before completion by default (see -Timeout)

Several minor fixes including:

• Disable Windows Recovery Console in fleet VMs
• Fix: Show-Fleet IOPS view now aggregates all VM disk devices
• Fix: clean up leaked/conflicting data collectors and blg automatically

VM Fleet 2.0 9/22/2021

• major release and rewrite as a first class Powershell module
• original script-based VM Fleet remains available at Frameworks/VMFleet1.0
• see the documentation in the Wiki

Source Code

The source code for DiskSpd is hosted on GitHub at:

https://github.com/Microsoft/diskspd

Any issues with DiskSpd can be reported using the following link:

https://github.com/Microsoft/diskspd/issues