Choosing a VPS host based on marketing copy is a waste of time. You need data.
I ran standardized tests across nine popular VPS providers to measure what matters: how fast instances boot, how responsive the network is, and whether disk I/O holds up under load. Some hosts surprised me. Others fell flat.
Why these metrics matter
Boot time tells you how quickly you can scale or recover from a crash. If your host takes three minutes to spin up a new instance, you're hemorrhaging uptime during an outage.
Network latency affects every HTTP request, database query, and API call. A provider routing traffic through congested peers will bottleneck your application no matter how much CPU you throw at it.
Disk I/O separates real SSD performance from oversold storage. When twenty VMs share the same physical volume, you'll see it in your database slow query log.
Test methodology
I deployed identical 2-vCPU, 4 GB RAM instances across all nine providers. Each ran Ubuntu 22.04 LTS with no custom kernel tweaks.
For boot time, I measured from API call to SSH-ready state using automated scripts. Five cold boots per host, median value recorded.
Network latency used ping and mtr to three geographically distributed targets over a 24-hour window. I captured mean, p50, and p99 values.
Disk I/O testing combined sequential and random workloads via fio with a 4 GB test file:
fio --name=seqwrite --rw=write --bs=1M --size=4G --numjobs=1 --runtime=60 --time_based
fio --name=randread --rw=randread --bs=4k --size=4G --numjobs=4 --runtime=60 --time_based
All tests ran three times. I discarded outliers more than two standard deviations from the mean.
The contenders
I tested DigitalOcean, Linode (Akamai), Vultr, Hetzner, OVHcloud, UpCloud, Contabo, Scaleway, and Kamatera. Each offers hourly billing and bare Ubuntu images.
No sponsored placements here. I paid retail prices for every instance.
Boot time results
Boot speed varied wildly—from under twenty seconds to nearly two minutes.
The fastest hosts used KVM with optimized block drivers and minimal launch overhead. UpCloud and Vultr consistently booted in eighteen to twenty-two seconds. Linode followed at twenty-eight seconds.
DigitalOcean landed in the middle pack around forty seconds. Not slow, but not impressive either given their infrastructure scale.
Contabo struggled. Median boot time hit ninety-four seconds, with one run taking over two minutes. That's unacceptable for auto-scaling or quick failover.
Hetzner and OVHcloud both clocked in near forty-five seconds—adequate but unremarkable.
What slows boot times?
Host-level resource contention is the usual culprit. When the hypervisor is oversubscribed, your VM waits in queue for CPU scheduling before it even starts POST.
Older SAN storage adds latency too. Providers still running spinning disks for boot volumes will never compete with local NVMe.
Some platforms inject excessive cloud-init scripts at launch. If the vendor's metadata service is slow, your instance hangs until the network stack times out.
Network latency and routing
Latency testing showed that datacenter location matters less than peering quality.
I measured round-trip times to three public targets: a major CDN edge, a cloud DNS resolver, and a well-connected looking glass server.
Vultr and UpCloud delivered the lowest mean latency—under two milliseconds to nearby targets and sub-fifteen milliseconds cross-country. Their p99 stayed tight, meaning consistent performance even during peak hours.
Linode's network performed well in most regions but showed occasional spikes in p99 latency. Nothing terrible, just less predictable than the top tier.
DigitalOcean's latency was acceptable but routing sometimes took inefficient paths. One of my test runs to the west coast bounced through an extra AS before reaching the destination.
Hetzner and Scaleway both offered solid European connectivity. Cross-Atlantic performance dropped off faster than US-based hosts, which is expected.
Contabo's network was the weak link. Mean latency sat twenty to thirty percent higher than competitors, and p99 numbers spiked unpredictably. Packet loss hovered near one percent—not catastrophic but noticeable in SSH sessions and HTTPS connections.
Check your own routing
Run mtr from your VPS to your primary traffic sources:
mtr -r -c 100 your-api-endpoint.com
Look for consistent hop counts and stable latency. If you see ten-plus hops or sudden jumps in round-trip time, the provider's peering is suboptimal.
Disk I/O performance
Sequential write speeds tell you how fast the underlying storage is. Random read IOPS tell you whether it's oversold.
UpCloud dominated here with NVMe-backed volumes delivering over two thousand MB/s sequential writes. Random read IOPS exceeded sixty thousand. That's local SSD performance, not shared SAN.
Vultr and Linode both hit respectable numbers—around eight hundred MB/s sequential and thirty thousand random read IOPS. Fast enough for most workloads including busy databases.
DigitalOcean's standard SSD volumes came in near six hundred MB/s sequential. Random reads were decent at twenty-five thousand IOPS, but I noticed more variance between test runs than on other platforms.
Hetzner surprised me. Their budget pricing led me to expect slower disks, but sequential writes hit seven hundred MB/s and random reads stayed consistent around twenty-eight thousand IOPS.
Contabo lagged significantly. Sequential writes barely cracked three hundred MB/s, and random read IOPS sat under twelve thousand. That's noticeably slow for database-heavy apps or high-traffic WordPress installs.
Why random IOPS matter more
Most web applications issue many small reads, not large sequential operations. Your database runs SELECT queries against indexes, your filesystem serves thousands of small cache files, and your logs write in short bursts.
If random IOPS drop below twenty thousand, you'll see it in query response times and page load speeds. A VPS that benchmarks fast on sequential writes but chokes on random reads is oversold.
Price-performance analysis
Raw performance means nothing if you're paying three times the market rate.
Hetzner offered the best value. Their instances cost half what DigitalOcean charges for similar specs, and performance held up across all three tests.
Contabo is cheap—sometimes extremely cheap—but you get what you pay for. Boot times drag, network quality suffers, and disk I/O can't keep pace. Fine for dev environments or personal projects, risky for production.
UpCloud delivered top-tier performance but charged premium prices. Worth it if you need guaranteed speed and have the budget. Not the best fit for cost-sensitive workloads.
Vultr and Linode both sit in the sweet spot: competitive pricing with strong performance across the board. Either one works well for most hosting use cases.
What about CPU performance?
I didn't benchmark CPU directly because most VPS workloads are I/O-bound, not CPU-bound. Your web server spends more time waiting on disk reads and database queries than crunching numbers.
That said, all tested providers used modern Intel Xeon or AMD EPYC processors with similar single-thread performance. The difference between hosts comes down to CPU steal time—how much your vCPUs get preempted by other tenants.
Check CPU steal with top or htop. If you consistently see steal above five percent, your host is oversubscribed.
Multi-region availability
Global reach matters if you serve international traffic or need failover across continents.
DigitalOcean, Vultr, and Linode all offer a dozen-plus datacenters spanning North America, Europe, Asia-Pacific, and increasingly South America and Africa.
Hetzner focuses on Europe with a few US locations. Excellent if your users are concentrated there, limiting if you need broad geographic distribution.
UpCloud and Scaleway provide solid European coverage with select US and Asian sites. OVHcloud has extensive European and North American presence.
Contabo's datacenter footprint is smaller—mostly Germany and US east/west. That limits your options for latency optimization.
Support quality and uptime
Performance tests don't capture everything. Support responsiveness and network stability matter when things break.
Linode and DigitalOcean both offer robust documentation, active community forums, and ticket response times under two hours for most issues. Their status pages are transparent about outages.
Vultr's support is hit-or-miss depending on your plan tier. Free-tier tickets can take a day or more. Documentation quality is decent but less comprehensive than the big two.
Hetzner's support is notoriously slow for anything beyond basic inquiries. If you need hand-holding, look elsewhere. Their docs are thorough, though, so experienced admins can usually self-serve.
UpCloud provides solid support with faster-than-average response times. Scaleway's support quality varies by region and language.
Contabo's support is... minimal. Expect long waits and generic responses. Fine if you're self-sufficient; frustrating if you're not.
Which VPS host should you choose?
For most production workloads, Vultr or Linode offer the best balance. Both deliver strong performance, competitive pricing, and reliable support.
If budget is tight and you're comfortable managing your own infrastructure, Hetzner is hard to beat. Performance holds up and the savings are real.
Need absolute top-tier speed and willing to pay for it? UpCloud consistently outperformed competitors on disk I/O.
Running dev environments or personal projects where occasional slowness is acceptable? Contabo will save you money at the cost of performance.
DigitalOcean remains a solid choice if you value ecosystem maturity—their marketplace, one-click apps, and integrations are extensive.
Frequently asked questions
Q: Can I trust synthetic benchmarks for real-world performance?
Benchmarks show potential, not guarantees. Your actual performance depends on your workload, time of day, and how many neighbors are hammering shared resources. Still, consistent benchmark results across multiple runs are a good indicator.
Q: Do managed VPS options perform differently?
Managed plans usually run on the same infrastructure as unmanaged. You're paying for patching, monitoring, and support—not faster hardware. The exception is dedicated CPU plans, which eliminate noisy neighbor problems.
Q: Should I care about IPv6 support?
Yes. IPv6 is increasingly required for certain compliance frameworks and improves routing efficiency. All tested providers except Contabo offer native IPv6 by default.
Q: How often should I re-benchmark my VPS?
Run basic I/O and network tests quarterly. If performance degrades significantly, open a ticket or consider migrating. Providers sometimes oversell nodes over time.
What to test on your own VPS
Benchmarks from a controlled test don't reflect your specific workload.
Once you've chosen a provider, run these checks:
- Boot a test instance and time it from API call to SSH login
- Use
fioto test disk I/O with your typical access patterns - Run
mtrto your most common traffic destinations - Monitor CPU steal percentage under load with
top - Check network throughput with
iperf3between your VPS and your office or primary CDN
Give the instance a few weeks of production-like traffic before committing to annual billing. Performance can degrade as the host adds more tenants to the same physical hardware.
Real-world testing beats marketing claims every time.