VPS Disk I/O Benchmark Comparison 2026 — NVMe vs SSD vs HDD

Every VPS comparison site will show you a table of IOPS numbers. Hostinger gets 65,000. RackNerd gets 20,000. Hostinger wins, article over, right?

Not quite. Raw IOPS numbers tell you almost nothing about how your application will actually perform. A provider with 50K IOPS and good I/O scheduling can feel faster than one with 65K and inconsistent latency spikes. Queue depth matters. Latency percentiles matter. Whether the provider oversubscribes their storage nodes matters more than the marketing sticker on the drive.

So this article does something different. Yes, there is an IOPS table. But after the raw numbers, I measured what actually changes when you put a WordPress site, a MySQL database, and a Docker build pipeline on each provider's storage. That is where the interesting gaps appear.

How I Tested (and Why Most fio Benchmarks Are Wrong)

A confession: two years ago I forgot --direct=1 on my first benchmark run. Without that flag, fio goes through the Linux page cache — you are benchmarking RAM, not disk. My "60,000 IOPS" on a $3.50 BuyVM instance should have been a red flag. Half the VPS benchmark articles online make the same mistake. If someone reports 100K+ random IOPS on a budget SSD, they did not use direct I/O.

Here is what I ran on every provider's cheapest plan, during three time windows (morning, afternoon, 2 AM EST) over two weeks in February-March 2026:

Why 4K block size? That is the default ext4 allocation unit, and MySQL InnoDB uses 16K pages (four 4K blocks). Testing with larger blocks inflates throughput numbers but misrepresents real application access patterns.

Why queue depth 32? Simulates a moderately busy application — roughly 20-40 concurrent WordPress users. Queue depth 1 results are less useful because almost no production workload is single-threaded.

Every test ran three times per time window. The numbers below are medians of nine runs.

Raw fio Results — All 13 Providers

Here is the table everyone came for. But read past it — the context below changes how you should interpret these numbers.

The spread is massive. Hostinger's 65K read IOPS is 3.25x faster than RackNerd's 20K. For context, in our CPU benchmark, the gap between #1 and #13 was only 1.57x. Storage is the single widest performance gap among VPS providers in 2026.

But here is what the table does not tell you: how much of that gap you will actually feel.

The Real-World Gap: Where IOPS Actually Matter

After running fio, I deployed the same WordPress site (developer theme, WooCommerce with 500 products, no page cache) on seven of the thirteen providers and measured three things that real people care about:

Now things get interesting.

The top 5 are closer than the raw IOPS suggest. Hostinger's 65K IOPS versus Vultr's 50K is a 30% gap on paper. But WordPress TTFB difference is only 31ms (187ms vs 218ms). Why? Because at these IOPS levels, the storage is no longer the bottleneck — PHP execution time and MySQL query parsing dominate the latency budget. Once you cross roughly 40,000 IOPS, diminishing returns set in hard for typical web workloads.

The real cliff is below 30K IOPS. Contabo and RackNerd are where disk speed becomes the actual bottleneck. The 312ms and 348ms TTFB numbers are not just "a bit slower" — they represent a fundamentally different user experience. Under load, these servers start queuing I/O requests, and latency spikes become unpredictable. I saw p99 latencies above 800ms on RackNerd during the 50-concurrent-user test. On Hostinger, p99 stayed below 320ms.

Docker builds reveal the gap most clearly. Building a Node.js image involves thousands of small file ops: extracting layers, running npm install, compiling assets. The 34s vs 79s gap between Hostinger and RackNerd is 2.3x — and if you run CI/CD on your VPS, that hits every single build. See our Docker VPS guide for the full breakdown.

NVMe vs SSD vs HDD — Cutting Through the Noise

Every VPS provider's marketing page says "SSD storage" like it is a feature. In 2026, SSD is not a feature. It is the floor. The question is what kind of SSD, how it connects to the CPU, and how many tenants share it.

NVMe: The Direct Line

NVMe plugs directly into the CPU through PCIe lanes — no SATA controller in the middle. Two structural advantages: parallelism (65,535 queues vs SATA's single queue of 32 commands) and lower command overhead (2.5 microseconds vs 6 microseconds per operation). When your database has 30 queries in flight, NVMe barely notices. SATA starts queuing. Hostinger and ScalaHosting — the only two NVMe providers — hold the top two positions. See our best NVMe VPS page for the full list.

SATA SSD: The Competent Middle

Most providers use SATA SSDs, with a ceiling of roughly 550 MB/s sequential and 55-60K random IOPS for enterprise drives. DigitalOcean hits near that ceiling (likely Samsung PM893 or Intel D3-S4520 enterprise drives). Contabo's 25K suggests consumer-grade or heavily shared storage. The label "SSD" tells you the media type but nothing about controller quality, RAID config, or oversubscription. DigitalOcean's SSDs are 2.2x faster than Contabo's. Same technology, completely different experience.

HDD: Dead for Primary Storage

No provider in our test set uses HDD for primary storage. Mechanical drives deliver 100-200 random IOPS — 300x slower than the worst SSD here. You will only encounter HDD in "storage VPS" add-ons. Fine for backups. Unacceptable for anything else.

The Metric Nobody Talks About: I/O Consistency

Here is where I diverge from every other VPS benchmark article. Raw IOPS medians are the polished number — the average day at work. What matters more is the worst day. Specifically: how much do IOPS fluctuate between test runs, and what happens at the tail of the latency distribution?

I ran the same fio read test nine times per provider (three times at three different times of day). Here is the coefficient of variation — lower is better:

This reshuffles rankings. Hetzner's median (52K) is lower than ScalaHosting's (58K), but Hetzner is almost twice as consistent (3.8% vs 6.1% CV). Users notice inconsistency more than average speed — Hetzner punches above its IOPS class.

Contabo and RackNerd's 18-21% CV means their 2 AM and 2 PM performance can differ by 40%+. That is oversubscribed shared storage — when neighbors run backups, your IOPS tanks. You cannot engineer around inconsistency with caching. You can only leave.

Ranked by Workload: Who Wins for What

Different workloads stress different aspects of storage. Here is how the rankings shift depending on what you are actually running:

For MySQL/PostgreSQL Databases

Database queries are random-read-heavy with burst writes on commits. The ranking that matters is: read IOPS first, write IOPS second, consistency third.

1. Hostinger — 65K reads, 55K writes, 3.2% CV. The full package.
2. Hetzner — 52K reads, 44K writes, 3.8% CV. Lower peak but ultra-consistent. A database that never has I/O spikes is worth more than one that is fast on average.
3. DigitalOcean — 55K reads, 42K writes, 5.1% CV. Strong across the board, plus managed database add-ons if you want to offload entirely.

For WordPress (Uncached)

WordPress without object caching hammers the database with 30-80 queries per page. Under concurrency, the mixed read/write pattern dominates. Our TTFB tests tell the real story:

1. Hostinger — 187ms TTFB at 50 concurrent users. The NVMe advantage shows under load.
2. DigitalOcean — 203ms. Only 16ms behind. The managed WordPress plan also pre-optimizes I/O settings.
3. Vultr — 218ms. Solid, and the Vultr High Frequency instances with NVMe would rank higher, but those start at $12/mo.

For the full WordPress performance picture, see our WordPress load test benchmark.

For Docker and CI/CD

Docker builds are sequential I/O beasts. Layer extraction, npm/pip installs, and asset compilation generate thousands of small file creates and reads in rapid succession. Write IOPS and write consistency are what matter here.

1. Hostinger — 34s build time, 55K write IOPS. NVMe's queue depth shines during parallel npm installs.
2. Hetzner — 41s build. The consistency story again: Hetzner's write IOPS (44K) actually beats DigitalOcean's (42K), and the lower variance prevents random slowdowns mid-build.
3. DigitalOcean — 38s build. Fastest SSD option, excellent for teams already using their container registry and Kubernetes offerings.

For Static Sites and Cached WordPress

If you are serving static HTML or using a full-page cache like WP Super Cache, your disk I/O barely matters. Once pages are in the OS page cache (RAM), subsequent reads never touch the disk. Even RackNerd's 20K IOPS is fine here. Save your money. Pick a provider based on network speed instead.

IOPS Per Dollar — The Value Equation

Absolute performance is half the picture. Here is how much random read IOPS you get per dollar of monthly spend, which reorders the rankings in surprising ways:

Hetzner takes the value crown, and also scored well in our individual benchmark. High IOPS, low price, best consistency. The catch: only one US datacenter (Ashburn, VA). If you need West Coast presence, Hetzner cannot help you.

Contabo is the real outlier. At $6.99/mo — more expensive than Hostinger — it delivers less than half the IOPS. Its value proposition is 200GB of storage space, not speed. ScalaHosting's last-place value ranking needs context: you are paying for managed hosting, SPanel, and security — not just disk speed.

When Tuning Beats Upgrading: I/O Optimizations That Actually Work

Before you migrate to a faster provider, consider this: I have seen a properly tuned Contabo VPS (25K IOPS) outperform an untuned Hostinger VPS (65K IOPS) for WordPress. Four changes make the biggest difference:

1. Set innodb_buffer_pool_size to 70-80% of RAM. Most providers ship MySQL with a 128M default. On a 2GB VPS, set it to 1400M. This keeps your working dataset in memory and eliminates most random reads.

2. Add Redis object caching. Reduces MySQL disk reads by 90% for repeat visitors. On a VPS below 35K IOPS, this is not optional.

3. Mount with noatime. Stops Linux from writing an "accessed" timestamp on every file read. Free IOPS, zero downside. Add it to /etc/fstab.

4. Move temp files to tmpfs. PHP sessions and application caches can live in RAM: tmpfs /tmp tmpfs defaults,noatime,size=256M 0 0 in fstab.

Combined, these make a 25K IOPS drive feel like 50K for web workloads. For providers in the 30-45K range, tuning often closes the gap entirely. See our VPS optimization guide for the full checklist.

The Verdict: What to Actually Buy

After spending three weeks testing, measuring, and writing this, here is what I would actually do with my own money for different use cases:

One thing I would not do: choose Contabo or RackNerd for anything that touches a database under load. The raw IOPS are low, the consistency is poor, and no amount of caching fully compensates for I/O variance above 15%. Those providers have their place — Contabo for bulk storage, RackNerd for ultra-budget static hosting — but disk-intensive workloads are not it.

FAQ — 7 Questions That Actually Matter

How much do IOPS actually matter for WordPress performance?

More than most people think, but less than IOPS-obsessed benchmark sites suggest. A typical WordPress page load triggers 30-80 MySQL queries. Each query involves multiple random 4K reads. On a 65K IOPS NVMe drive, those reads complete in under 2ms total. On a 20K IOPS budget SSD, it is closer to 6ms. The difference — about 4ms — sounds trivial. But under 50+ concurrent users, those milliseconds multiply. We measured a 120ms TTFB difference between Hostinger (NVMe) and RackNerd (budget SSD) serving the same WordPress site under load. So IOPS matter, but only under concurrency. A single visitor will not notice.

Is NVMe actually faster than SSD for VPS hosting, or is it marketing?

It is real, but the gap is smaller than spec sheets suggest. NVMe theoretical max is 3,500-7,000 MB/s versus SATA SSD's 550 MB/s cap. In practice, our fio benchmarks showed NVMe providers averaging 15-25% higher random IOPS than the best SSD providers. DigitalOcean's enterprise SSDs hit 55K IOPS — only 15% behind Hostinger's NVMe at 65K. The biggest NVMe advantage is latency under queue depth: when 32+ operations are queued simultaneously, NVMe's 64K queue depth versus SATA's 32-entry limit becomes the bottleneck. For database-heavy workloads with high concurrency, NVMe is measurably faster. For a blog with caching, you would never notice.

Why did Contabo score so poorly despite having 200GB of storage?

Because storage capacity and storage speed are completely different metrics. Contabo's business model prioritizes giving you the most gigabytes per dollar. They achieve this with high-density storage arrays that pack more drives per node, which introduces contention. Our fio tests showed 25K read IOPS — adequate for light workloads but 2.6x slower than Hostinger. Contabo also showed the highest variance between test runs (±18%), suggesting noisy neighbor effects on shared storage. If you need bulk storage for media files or backups, Contabo is fine. If you need fast random I/O for databases, look elsewhere.

Can I improve disk I/O on my existing VPS without switching providers?

Yes, significantly. Three approaches work: (1) Move hot data to RAM — use Redis for session storage, OPcache for compiled PHP, and tune MySQL's innodb_buffer_pool_size to 70-80% of available RAM. This eliminates most random reads. (2) Reduce write amplification — switch MySQL to innodb_flush_log_at_trx_commit=2 (slight durability trade-off for 10-50x write speed improvement), disable access logging for static assets, and use tmpfs for PHP session files. (3) Use filesystem optimizations — mount with noatime, use ext4 with journal_data_writeback mode. These changes can make a 25K IOPS drive feel like 50K for most web workloads.

How do I test disk I/O on my own VPS?

Install fio (apt install fio on Ubuntu/Debian) and run: fio --name=randread --ioengine=libaio --direct=1 --bs=4k --iodepth=32 --size=1G --rw=randread --runtime=60 for read IOPS. Replace randread with randwrite for writes. Critical: use --direct=1 to bypass page cache (otherwise you are benchmarking RAM, not disk). Run each test 3 times and take the median. Run at different times of day — some providers show 20-30% IOPS drops during peak hours due to shared storage contention. If your results are more than 25% below our numbers for the same provider, open a support ticket. You may be on an oversubscribed storage node.

Does RAID configuration on VPS hosts affect my IOPS?

Absolutely, and it is one of the hidden variables in VPS benchmarking. Most providers use RAID-10 (mirrored stripes), which doubles write latency because every write must hit two drives. Some use RAID-5/6 with parity, which is even worse for writes — that is likely why certain providers show read-to-write ratios above 1.3:1. The provider's storage controller, caching policy, and oversubscription ratio matter more than whether the drives are labeled "NVMe" or "SSD." We have seen "NVMe" providers with worse IOPS than good SSD setups because the storage array behind the label was poorly configured.

Should I choose a provider based on IOPS alone?

No. Disk I/O is one of four performance pillars alongside CPU, RAM, and network speed. A VPS with 65K IOPS but a weak single-core CPU will still render PHP pages slowly. That said, disk I/O is the metric most likely to cause sudden performance cliffs — everything else degrades gradually under load, but when disk I/O saturates, response times spike exponentially because the kernel's I/O scheduler starts queuing requests. Our recommendation: first filter by CPU performance, then use disk I/O as the tiebreaker. For database-heavy workloads specifically, flip that priority.