NVMe vs SATA SSD on a VPS: benchmarks from our own fleet

Test methodology

Run on production nodes in Frankfurt during a low-traffic window, from inside a 4 vCPU / 8 GB guest, so these are numbers a customer could reproduce — not host-level figures with the guest overhead removed.

fio --name=randread --ioengine=libaio --direct=1 --bs=4k \
    --iodepth=32 --size=4G --runtime=60 --numjobs=4 --group_reporting --rw=randread

4K random read, QD32

* Capped by the provider's per-volume IOPS limit, not by the hardware.

4K random write, QD32

What that means in workload terms

• PostgreSQL, 8 GB dataset, pgbench 50 clients: 4,180 tps on NVMe vs 940 tps on SATA. The database is latency-bound on WAL flushes, which is exactly NVMe's strength.
• Windows Server boot to usable desktop: 19 s vs 41 s. Not important once, very noticeable when you reinstall or reboot regularly.
• npm install on a large monorepo: 47 s vs 112 s. Package managers are a flood of tiny random writes.
• Serving static files from nginx: essentially identical. Once the working set is in page cache, the disk is irrelevant — do not pay for NVMe to solve a problem you do not have.

Why "1,000,000 IOPS" in an ad means nothing

Three reasons that number is usually theatre:

1. It is the node's total, not yours. A node hosting 40 guests shares that pool. Your share depends entirely on what the neighbours are doing.
2. It is a burst figure. Consumer-class NVMe uses an SLC cache; once it is full, sustained write speed can fall by 70%. Enterprise drives with power-loss protection hold their numbers, and cost about 3x more per terabyte.
3. Queue depth is unstated. QD256 with 16 jobs produces a big number that no real application generates.

The honest metric is p99 latency under a realistic queue depth, sustained for at least ten minutes. That is why the tables above include p99.

Is your disk actually the bottleneck?

Linux

iostat -xz 1        # %util near 100 with high await = disk-bound
vmstat 1            # high 'wa' in the CPU columns = waiting on I/O
iotop -oPa          # which process is responsible

Windows

Get-Counter '\PhysicalDisk(_Total)\Avg. Disk sec/Read' -Continuous
# Sustained above 0.020 s (20 ms) means the disk is the constraint.

If await is low and %util is low but the app is still slow, the disk is innocent — look at CPU steal, memory pressure or a network round trip.

What we run

Every RDPForge node uses enterprise NVMe (Samsung PM9A3 / Micron 7450 class) in RAID-10 with power-loss protection and hardware-accelerated parity. There is no SATA tier left in the fleet — the last SATA node was retired in November 2023 — and there is no network-attached block storage anywhere, so your disk is always physically in the machine running your VM.