Microsoft has begun shipping Windows Server 2025, offering enhanced security, improved performance, and cloud flexibility. The new OS is designed to meet the demands of today’s most demanding workloads, including AI and machine learning. In particular, Radically improved NVMe storage performance is claimed : Windows Server 2025 delivers up to 60% greater storage performance in IOPs compared to Windows Server 2022 on identical systems.
According to According to Microsoft, the upgraded NVMe support includes:
• Redesigned low-level Windows storage stack to optimize high-performance multi-queue storage and take advantage of NVMe technology features (StorMQ)
• Extended Windows Storport driver model to support existing miniports and new StorMQ miniports
• A new architecture that significantly improves CPU utilization, performance, and scalability compared to the existing stack.
Checks the mo personally
Storages are created using Storage Spaces, a Microsoft technology for managing data storage spaces. In detail – here .
The choice of media significantly affects server performance. NVMe SSDs have the highest potential among all types of storage, thanks to their direct connection to the PCIe bus and central processors, without intermediate controllers. NVMe SSDs will always be faster than SAS and SATA SSDs, despite the fact that they are all made on the same NAND memory. They are also currently the cheapest.
Let's examine the impact of OS choice on Storage Spaces performance on NVMe SSDs. The 60% gain from a rewritten software stack seems overly optimistic.
Our test bench configuration:
Server DELL R7615 / 1 x AMD EPYC 9174F / 4 x 32GB DDR4-4800 RDIMM / 2 x 480GB M.2 NVMe SSD / 2 x 1.6TB U.3 / Windows Server Standard 2022 + Windows Server Standard 2025.
Let's limit ourselves to a few tests of a two-way mirror of NVMe, the simplest of protected storage spaces.
One of the bootable NVMe drives was installed with an instance of Windows Server 2022 (21H2, build 20348.2849), and the other was installed with Windows Server 2025 (24H2, build 26100.2314).
What was measured, what was measured?
To create a load on the disk subsystem and measure its performance, Microsoft recommends using the DiskSpd utility (https://aka.ms/diskspd). It can perform I/O operations in multiple threads with a specified target. The utility does not require installation; for x64-bit systems, we use the diskspd.exe version from the amd64 directory.
Command to test the disk:
diskspd.exe -c50G -d300 –t16 -o32 -b4k -r -w30 -Sh -LE:io > testdata.txt
- c50G – generate a test file of size 50GB;
• -d300 – test duration in seconds;
• -t16 – number of threads (taken equal to the number of CPU cores);
• -o32 – queue length;
• -b4k – block size (we use values of 4K and 64K);
• -r – random read/write;
• -w30 – ratio of write operations to read operations 30%/70%;
• -Sh - do not use caching.
The DISKSPD test provides results that are comparable to real-world workloads. However, it is important to pay attention to the parameters specified and their relevance to the real-world scenario. It is important to understand that artificial workloads will never perfectly represent the real-world workload of an application during deployment.
We measured performance in IOPS using three tests: 4K short block write, database test (70% read, 30% short block write), database test (70% read, 30% 64K block write). Results in the charts.
Result
There is a performance gain, but it is far from the declared one. Hopefully, with the release of updated drivers, the difference will be greater. In any case, new servers should be deployed with a new OS, so the comparison is more of an academic interest. It is important that Microsoft pays special attention to the performance of NVMe storage, everything is going to the point that there will soon be no other media in servers.