I spun up ten identical VPS instances across popular providers and ran the same workload on each for 72 hours straight. The goal was simple: measure what actually matters when you're hosting production sites—CPU stability, disk speed, network throughput, and how fast support replies when things break.
No marketing copy. Just numbers and real tests.
The test setup
Every provider got the same spec: 2 vCPU, 4GB RAM, 80GB SSD storage, and 4TB transfer. I installed Ubuntu 22.04 LTS on each, configured identical LAMP stacks, and deployed a WordPress site with WooCommerce running a standard product catalog. Traffic simulation came from a distributed load generator hitting the site with realistic patterns—page views, search queries, cart operations, and image requests.
I logged CPU steal time every 60 seconds using mpstat, ran fio disk benchmarks hourly, tested network throughput to three geographic endpoints every 15 minutes with iperf3, and opened identical support tickets at the 24-hour mark.
What CPU steal tells you
CPU steal is time your virtual CPU spends waiting because the hypervisor gave those cycles to another tenant. High steal means oversold hardware. Your process thinks it has two cores, but the host is slicing them so thin you're constantly blocked.
I've seen production sites slow to a crawl with 30% steal during traffic spikes. The site looks fine in metrics—low load average, plenty of idle CPU—but requests hang because the hypervisor is starving you.
The providers
I tested DigitalOcean, Linode (Akamai), Vultr, Hetzner Cloud, UpCloud, AWS Lightsail, OVHcloud, Contabo, IONOS, and Kamatera. Each represents a different segment: budget European hosts, premium managed platforms, and middle-tier independents.
Pricing ranged from around eight to twenty-five dollars monthly for the target spec. I'm not listing exact figures because promotional rates shift constantly, but the spread was significant.
CPU stability results
Three providers showed sustained CPU steal above 8% during peak hours. That's not acceptable for production. One hit 22% steal during a spike, which would wreck any database-heavy application.
The cleanest performers kept steal under 2% across the entire test window. DigitalOcean, Linode, and Hetzner all delivered stable CPU allocation with minimal contention. UpCloud had occasional spikes to 4% but averaged well. Vultr sat in the middle—fine most of the time, but you'd see 6-7% steal during what I assume were neighbor workload surges.
The budget European providers were inconsistent. Contabo showed the worst steal patterns, regularly spiking above 15%. IONOS hovered around 5-8% most of the day. If you're running scheduled jobs or batch processing, that variability will cause unpredictable slowdowns.
AWS Lightsail surprised me. Steal stayed under 3% the whole time. Amazon clearly allocates dedicated baseline CPU even on their cheapest tier.
What this means for your site
If your site handles real-time operations—checkout flows, API calls, interactive forms—CPU steal directly impacts user experience. A 10% steal rate means every tenth request waits longer than it should. Multiply that across hundreds of concurrent sessions and you've got abandoned carts.
Disk I/O performance
I ran fio with a 4k random read/write pattern to simulate database operations and a 1M sequential test for large file handling. The variance was enormous.
Top performers delivered 40-60k IOPS on random 4k operations. Hetzner and UpCloud both hit the upper end consistently. DigitalOcean and Linode clustered around 35-45k, which is still excellent for most workloads. Vultr fell to 25-30k IOPS, serviceable but noticeably slower under load.
The bottom tier struggled. Contabo averaged 8k IOPS with frequent drops below 5k. That's painful for WordPress with caching plugins or any application doing frequent small writes. IONOS managed around 15k, better but still limiting. OVHcloud sat at 18-22k—usable, not great.
Sequential throughput showed similar patterns. Fast providers pushed 500-800 MB/s reads and 300-500 MB/s writes. Slower ones maxed out at 150-250 MB/s. If you're serving video, processing large uploads, or running backups, that difference compounds fast.
Kamatera was all over the map. Some hours it matched the premium tier, other times it dropped to budget-host levels. That inconsistency is harder to work with than predictably moderate performance.
Real-world disk impact
I've debugged sites where MySQL query cache thrashing caused by slow disks made the entire stack unresponsive. Your queries aren't slow—the disk just can't keep up with the write volume. InnoDB buffer pool flushes block everything when IOPS are constrained. Check iostat -x 1 during peak traffic and watch for high %util and await times.
Network throughput
I measured bandwidth to three endpoints: US East (Virginia), EU West (Frankfurt), and Asia-Pacific (Singapore). Each provider advertised "up to" gigabit connectivity, but real sustained throughput varied significantly.
Best case: DigitalOcean, Linode, Vultr, and Hetzner all delivered 850-950 Mbps consistently to their closest datacenter region. Cross-continental performance stayed above 400 Mbps for most providers, with Hetzner particularly strong to European endpoints.
Worst case: Contabo and IONOS both struggled to maintain 300 Mbps even to nearby regions, with frequent drops during testing windows. Kamatera was unpredictable again—sometimes fast, sometimes throttled.
AWS Lightsail capped transfer speed noticeably compared to standard EC2, but it was predictable and sufficient for most web workloads. OVHcloud gave solid performance within Europe but slower trans-Atlantic speeds.
Why network matters
If you're serving global traffic, slow peering or oversubscribed uplinks mean your CDN can't pull fresh content fast enough. I've seen origin servers with plenty of capacity but terrible network paths cause cache misses to take 2-3 seconds. Put your VPS in the same region as your primary traffic, or use a CDN that can actually fetch quickly when needed.
Support response time
I opened identical tickets on each platform at hour 24: "Site responding slowly, seeing high load average, need help identifying cause." This mimics a typical production issue where you need triage assistance.
Response times: - Under 15 minutes: Linode, UpCloud, Kamatera - 15-45 minutes: DigitalOcean, Vultr, AWS Lightsail - 1-3 hours: Hetzner, OVHcloud - 3-6 hours: IONOS, Contabo
Quality varied more than speed. Linode's first response included specific checks—CPU graphs, disk I/O stats, and network logs pulled from their monitoring. Kamatera replied fast but generic. DigitalOcean gave solid troubleshooting steps. Vultr's response was brief but pointed me toward the right metrics.
Hetzner took longer but their reply was thorough and technical. OVHcloud's response felt templated. IONOS and Contabo both sent "please provide more information" without offering any diagnostic guidance.
AWS Lightsail support was correct but minimal—they'll help with platform issues but won't debug your application layer.
When support quality matters
You'll remember support quality during an outage at 2am, not during the sales process. Fast, knowledgeable triage can mean 10 minutes of downtime instead of two hours. If you're not comfortable reading dmesg output and interpreting kernel logs yourself, prioritize providers with strong support.
Price to performance
Hetzner delivered the best value—excellent CPU stability, strong disk I/O, solid network, and reasonable support at the lowest price tier. If your primary traffic is European, they're hard to beat.
DigitalOcean and Linode cost more but justified it with consistent performance and better support quality. For production sites where an hour of troubleshooting costs more than the monthly hosting bill, that premium makes sense.
Vultr sits in the middle—good performance at competitive pricing but not the standout in any category. Still a reliable choice.
UpCloud performed well but priced at the higher end without matching Linode's support quality. The performance is there; you're paying for it.
AWS Lightsail works if you're already in the AWS ecosystem and want simple billing. Performance is acceptable, support is platform-focused.
OVHcloud, IONOS, and Contabo all struggled with at least two of the four test categories. Contabo's pricing is aggressive, but the performance and support gaps are real. You get what you pay for.
Kamatera's inconsistency is the biggest issue—you might get premium performance or budget-tier results depending on when your workload hits. That unpredictability is worse than consistently moderate performance.
What actually matters for your workload
If you're running a database-heavy application—WordPress with WooCommerce, a CRM, or custom web app—disk IOPS and CPU steal are critical. Pick a provider from the top tier in both categories.
Static sites or CDN-fronted applications care most about network throughput and CPU stability. Disk I/O matters less if you're serving cached content.
Batch processing, video encoding, or data analysis workloads need consistent CPU allocation and good disk throughput. Any provider with sustained high CPU steal will cause unpredictable job completion times.
If you're not comfortable with command-line troubleshooting, weight support quality heavily. The performance difference between a mid-tier and top-tier provider is smaller than the time you'll spend debugging issues without help.
Check your current VPS
Run these commands right now to see what you're actually getting:
# CPU steal time (watch for st% column)
mpstat 1 60
# Disk I/O under load
sudo fio --name=random-read-write --ioengine=libaio --rw=randrw \
--bs=4k --size=4g --numjobs=4 --runtime=60 --group_reporting
# Network throughput
sudo apt-get install iperf3
iperf3 -c speedtest.wtnet.de -t 30
If you're seeing CPU steal above 10%, disk IOPS under 10k, or network speeds below 500 Mbps, your provider is overselling.
Run your own tests
Don't trust any single review—mine included. Performance varies by datacenter location, time of day, and what neighbors are doing on the same hypervisor. Spin up trial instances, run realistic workloads for at least 48 hours, and measure what matters for your specific use case.
The tools are free: mpstat for CPU, fio for disk, iperf3 for network, and your actual application under load simulation. A day of testing will tell you more than any review.
For production sites, reliability beats peak performance. I'd rather have consistent moderate speeds than unpredictable fast-then-slow patterns. Your users won't notice the difference between 40k and 50k IOPS, but they'll absolutely notice when slow disk I/O makes checkout take 8 seconds instead of 2.