All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
0% steal, 28% performance — my Contabo VPS 6 mystery, solved in 3.5 hours
tiana93046
Member
Nothing here is a revelation — Contabo overselling is well documented on this forum, and "support moved me to another host" stories exist too; the ones I found usually took days of template replies first (one reviewer's CPU ticket took five days of back-and-forth before they acknowledged a host issue). Sharing mine as a data point because two things were unusual: the timeline (Saturday evening, ~3.5 hours from ticket to both migrations done), and the failure mode — see the steal numbers below.
Setup
I already had a Cloud VPS 4 (4 vCPU / 8 GB, EU) running fine. On Nov 14 I ordered two Cloud VPS 6 (6 vCPU / 12 GB, EU, $9/mo each) and benchmarked them against the VPS 4 before putting them into production.
The problem
Single-thread sysbench (prime=20000): the VPS 4 does ~1,820 events/s. The two new boxes did 512–548 — about 28% of the old box. All-core throughput was 36% and 46% of the VPS 4. OpenSSL RSA-2048 sign: 0.53 ms on the old box vs 1.07 ms and 0.80 ms on the new ones. Same price class, a third of the CPU. Odd part: steal was 0.000% on both new boxes the whole time. Most slow-Contabo reports blame steal; whatever was wrong with these hosts, it wasn't steal — the cores were simply slow.
What I sent support
Not "server slow pls fix". A measurement dossier, in three messages over one evening:
- Three benchmark families (sysbench single-thread and all-core, Python loop, OpenSSL sign/verify + AES-GCM)
- Controls: identical sysbench binary on all three machines (SHA-256 matched), steal at 0.000% on all of them, no cgroup throttling, new boxes actually consuming their full 6 cores — so the gap had to be host-side
- An equal-thread comparison (4 threads everywhere) to rule out core-count effects
- A healthy in-account reference machine (the VPS 4) as the baseline
- A concrete ask: check the two host systems, and if needed migrate — I even proposed migrating one box first as an A/B test
What happened
Ticket filed Saturday evening (CET) through the panel. First human reply in ~3.5 hours: they had analysed both VPS and the hosts and decided to live-migrate both. Both migrations finished within the hour. IP addresses, data and root password unchanged; only the VNC endpoints changed.
Retest (OpenSSL RSA-2048 sign, single thread, median of 3)
- Box A: 934 → 2,387 signs/s (2.55× — now above the old VPS 4's 1,794)
- Box B: 1,250 → 1,645 signs/s (1.32× — ~92% of the reference)
I replied with the retest numbers, confirmed the issue resolved, and the ticket could be closed.
Takeaway
Contabo's terms only guarantee connectivity (99.9%), not CPU — so a performance ticket is a negotiation, not a claim. "My server is slow" gets you the "everything is normal on our side" loop. A controlled benchmark set with a reference machine leaves support nothing to argue with, and migration is clearly a remedy they're willing to use when the data is there.
Anyone else seen a 0%-steal slow host, or is this a different failure mode? Happy to post the exact benchmark commands if anyone wants to replicate the test setup.
Comments
Thanks, chat gee pee tea
Contabo is part of the past for most of us.
ever since i got migrated to a new node one day ive also seen 0% steal completely, i am pretty sure they just ended up disabling steal reporting on their hosts so nobody can complain about how high the steal is
Steal is so high it rolled over to 0%
0% steal with slow cores usually means the contention isn't at the scheduler level, so steal can't see it. Common causes are your vCPUs landing on SMT siblings of a busy neighbour's threads, the node running an older CPU generation, or the host being power/thermal capped. Quick check is to compare
grep 'model name' /proc/cpuinfoand the CPU flags on all three boxes before and after the migration; if the model changed, that explains most of the 2.5x jump. Benchmarking in the first hour after provisioning is a good habit with any budget host, since that's when a node move is easiest to get.Ha, fair point — English isn't my first language, so I let AI help me turn my notes into readable English before posting. The benchmarks are all mine though, ran them on my own boxes.
that's an interesting theory. my old VPS 4 showed 0.000% steal the whole time too and it performed fine, so if they turned steal reporting off it's not just the bad hosts, seems to be everywhere. either way steal told me nothing here — the raw benchmark numbers were the only thing that mattered, that's why I built the ticket around them.
Contabo has been known for their overselling for many years now, it's nothing new