Burstable cores, and the arithmetic underneath them
A burstable vCPU is a promise about your average, not about your peak. For a great many jobs that is fine, and it is cheap. For anything with a hot thread and a deadline it is the most expensive line on the invoice, and it does not look like it.
What burstable actually sells
Written out plainly, without the language around it, a burstable plan is a subscription to a fraction of a core with an overdraft facility attached.
The model is perfectly reasonable. It is the specification sheet that is coy about it.
A baseline, not a core
You are entitled to a percentage of one physical core, commonly somewhere between five and forty, with permission to exceed it while credits last. The core count on the invoice is the ceiling, never the floor.
Credits accrue while you sleep
Idle time earns the overdraft that busy time spends. An instance that is genuinely busy for two hours a day works beautifully. One that is busy for six hits the baseline and stays there.
The ratio is the product
Overselling here is not fraud, it is the business model, and it works because most tenants are asleep at any given moment. Eight to twenty tenants per physical core is ordinary; the number is rarely published.
Contention is invisible from the price
Two instances with identical specification sheets, one on a quiet node and one on a busy one, differ by a factor you cannot see until you put load on them. You do not choose the node.
The silicon is usually older
Dense, cheap tenancy rewards amortised hardware. A burstable fleet is often two or three generations behind, which costs you again on every single-threaded operation.
How to measure it from inside the instance
None of this requires cooperation from the provider. Four measurements, one evening, on the cheapest plan they sell.
- 01
Read steal time under load
Both top and vmstat report st. On a genuinely dedicated core it sits at zero. Sustained figures above a percent or two mean the hypervisor is running somebody else on the silicon you are paying for.
- 02
Run for ten minutes, not for ten seconds
A short benchmark measures your credit balance. Loop a single-threaded integer benchmark for ten minutes and watch the score decay; wherever it flattens out is your real allocation.
- 03
Check the clock, then check it again
Record boost frequency at the start of the run and after five minutes at full load. A drop of forty percent is the node’s power and thermal budget being shared out among everyone on it.
- 04
Repeat at an inconvenient hour
Run the same test at 03:00 and again at 21:00 local to the site. A dedicated platform returns the same number twice. A contended one returns your neighbours’ schedule.
If a provider’s support desk cannot tell you the allocation ratio, that is itself the answer, and it took thirty seconds to get.
Two models, side by side
The left column is not a criticism, it is a description. Plenty of good operators run it deliberately and price it accordingly.
| Oversold burstable | Dedicated, as we sell it | |
|---|---|---|
| Core allocation | A share of a physical core, exceeded on credit | One physical core per core, allocated once |
| Tenants per core | Commonly eight to twenty | One |
| Steal time under load | Rises with node load, unknowable in advance | Zero, by construction |
| Sustained single-thread | Decays as the credit balance drains | Flat for as long as you care to run it |
| Silicon | Whatever fills the rack, frequently two generations old | Ryzen 9 9950X or EPYC 9354P, current generation only |
| Memory | Non-ECC on desktop boards, oversubscribed | DDR5, registered ECC on the EPYC line |
| Entry price | Three to eight euro a month is normal | €29 for four Zen 5 cores and 16 GB |
| Predictability | Depends on your neighbours’ week | Depends on your own workload |
| Where it hurts | p99 latency, build times, anything with a deadline | The invoice |
When a burstable host is the better buy
Paying dedicated prices for an idle machine is not prudence, it is waste with better branding. Several categories of workload should never come near us.
We do not sell anything below €29 a month, so for a whole class of jobs we are the wrong shop.
Duty cycle under five percent
A secondary DNS server, a monitoring probe, a jump host, a personal mail relay. These things idle for twenty-three hours and burst for a second, which is precisely what the burstable model is priced for.
Wide, cheap fan-out
Thirty small nodes in thirty cities to measure something from thirty angles. Buying dedicated cores thirty times over, to run a job that is asleep ninety-nine percent of the time, would be an odd way to spend money.
Learning and disposable work
Somewhere to break things, reinstall weekly and delete without regret. The refund policy here covers seven days; a burstable provider will sell you a month for the price of a sandwich.
A hard budget ceiling
If the real choice is a burstable instance or no instance at all, take the burstable one. It is a genuine computer, it will run your project, and it is a better outcome than not shipping.
What we do instead
Allocated once, counted once
A core you buy is a core nothing else is scheduled on. No burst credits, no steal time, no scheduler weighting, and no clause in the terms that quietly reverses this on a busy node.
Current silicon only
Ryzen 9 9950X for single-threaded work and EPYC 9354P where core count and ECC matter. Nothing older than the current generation is in production anywhere in the fleet, at any of the 34 sites.
Limits written down
The fair-use traffic figure is published per port rather than hinted at. Cross it consistently and you get an email about a dedicated 40 Gbit/s port, not a surprise invoice.
“The build that took nineteen minutes on the old box takes four here. I spent a week convinced I had misconfigured something.”
Questions
One. Cores, memory and NVMe are allocated at provisioning time and are not counted twice, which is the entire reason our entry plan starts at €29 rather than at €4.
Briefly, during a live migration, when a node is being drained for maintenance. That is a few seconds and it is announced in advance. Outside of that it should read zero, and a ticket showing otherwise gets investigated rather than explained away.
Yes, on every plan, for as long as you like. There is no throttle, no credit balance and no letter about unusual usage patterns. The cooling budget for a full node is already paid for.
No, and we are not planning to. The sensible move is consolidation: several idle jobs on one R-8 at €54 usually costs less than the small instances they replace, and you patch one operating system rather than six.
Four Zen 5 cores, and no queue in front of them
R-4 is €29 a month with 16 GB of DDR5 and 200 GB of Gen4 NVMe. Load it to the ceiling for a week, watch steal time stay at zero, and keep it only if the difference is real.