Virtual dedicated servers (VDS) are sold on vCPU counts and RAM megabytes, but the numbers on the invoice rarely describe what you actually get. Two plans can both advertise “8 vCPU / 32 GB” and deliver wildly different throughput once a noisy neighbour, a hypervisor steal cycle, or a NUMA boundary gets involved. If you are buying a VDS for a game backend, a busy database, or a high-traffic app, the only comparable unit is cost per usable core — the cores you can actually keep busy at the clock speed you pay for.
This guide takes three common VDS configurations and works through steal time, memory bandwidth, and the resulting cost per usable core. The methodology applies to any dedicated server hosting lineup, including bare-metal options when a VDS stops being the right answer.
Why “vCPU” Is Not a Unit of Work
A vCPU is a scheduling entitlement, not a physical core. Depending on the plan it can be:
- A pinned core — the vCPU is bound to a specific physical core (or SMT thread) and is never shared. You get the full clock plus any turbo that fits the thermal budget.
- A dedicated thread — bound to one SMT sibling, sharing the physical core’s execution units with another customer’s thread. Expect 10–25% throughput loss versus a pinned full core.
- An overshared vCPU — the hypervisor places your vCPU on any available core, possibly sharing it with several other tenants. On an idle host this feels fine; at peak it collapses.
Most budget VDS plans are the third type sold as if it were the first. The tell is in your own metrics, not the marketing page.
Measuring Steal Time: The Only Honest Test
Steal time is the percentage of time your vCPU was runnable but the hypervisor scheduled someone else. On a dedicated or pinned vCPU it stays below 0.2%. Run this for a week before you commit to long billing terms:
- Sample
/proc/statevery 10 seconds and record thestealfield per CPU:vmstat 10 360 | awk '{print $17}'. - Under a 50% sustained CPU load, a healthy pinned VDS shows p95 steal under 0.5%. An overshared plan shows 3–15% p95, and 20%+ during neighbour peaks.
- Cross-check with wall-clock drift against NTP: high steal shows up as clock offset growth inside the guest.
Steal below 1% means you can plan against the advertised specs. Steal above 3% means every capacity calculation you do is fiction — half your “8 vCPU” may not be schedulable when you need it.
Memory Bandwidth and the NUMA Trap
On a modern EPYC or Xeon host, a single physical core pulls 8–15 GB/s from local DRAM; a cross-socket fetch drops to 4–6 GB/s and adds 60–120 ns. A VDS with 8 vCPUs spread across two NUMA nodes behaves like two 4-core machines joined by a slow interconnect — fatal for a shared data structure such as a database buffer pool or a JVM heap.
| Placement | Effective mem BW (per core) | Added latency | Practical throughput |
|---|---|---|---|
| 1 CCD, local DRAM | 11–15 GB/s | 0 ns | 100% |
| Cross-CCD, same socket | 7–9 GB/s | 40–80 ns | 80–90% |
| Cross-socket | 4–6 GB/s | 90–140 ns | 60–75% |
| SMT sibling contention | — | — | 75–90% |
Check numactl --hardware inside the guest. A single-node layout with all vCPUs on one node is what you want. If the guest reports two nodes and memory is interleaved, your workload is paying a hidden tax on every cross-thread pointer chase.
Benchmark Three Plans on the Same Workload
A worked comparison of three 2026 VDS profiles on a 30-minute mixed workload: sysbench cpu --threads=N, 7-Zip LZMA, and an Nginx TLS loop pinned to one core. The final column is the number to compare against the invoice. For the equivalent analysis against physical machines, see bare metal vs virtualised latency benchmarks.
| Plan | Advertised | p95 steal | Bench score | Usable cores | $/mo | $/usable core |
|---|---|---|---|---|---|---|
| A: overshared, dual-socket | 8 vCPU / 32 GB | 11% | 0.62× | 4.4 | $60 | $13.60 |
| B: pinned threads, 1 CCD | 8 vCPU / 32 GB | 0.3% | 0.93× | 7.4 | $110 | $14.90 |
| C: full-core pinning, 3.8 GHz+ | 8 cores / 32 GB | 0.1% | 1.00× | 8.0 | $150 | $18.75 |
Plan A looks 40% cheaper per month. Per unit of work it is actually the most expensive, because 3.6 of its eight advertised cores do not exist when you need them. Plan B is generally the sweet spot for web services; Plan C only wins when your workload is tightly latency-bound (databases, real-time game servers, JVM heaps) and the extra single-thread headroom converts directly into revenue.
How to Model Cost per Usable Core
The usable-core model has three inputs and one output:
- Steal-adjusted cores = advertised vCPU × (1 − p95 steal).
- Placement penalty = memory/interconnect discount from the table above (0.75 cross-socket, 0.85 cross-CCD, 1.0 same-CCD).
- SMT factor = 0.85 if your workload is compute-dense and shares SMT siblings, 1.0 if pinned to physical cores.
- Usable cores = steal-adjusted × placement × SMT, divided by the benchmark-relative clock factor.
Whichever plan produces the lowest dollars per usable core and clears your latency requirements wins. Re-run the model at renewal — noise neighbours change, and a host’s fleet ages.
When per-Usable-Core Economics Favours Bare Metal
Above roughly 16 truly busy cores, the VDS premium stops buying anything: a pinned 16-core VDS at $18–22 per usable core sits around $290–350/mo, while a mid-range dedicated machine with 16–24 cores, 64–128 GB, and NVMe can be had for a comparable or lower price with zero steal time and no NUMA gamble. The crossover depends on your workload, and is worth modelling explicitly before you sign another annual contract.
Checklist Before You Sign
- Ask explicitly whether vCPUs are pinned, SMT-shared, or oversubscribed — in writing.
- Ask for the CPU model and host generation; a v3 Xeon and an EPYC 9754 both read “8 vCPU”.
- Run a 72-hour steal test before committing to 12-month billing.
- Verify NUMA topology inside the guest with
numactl. - Confirm memory is committed, not ballooned against host overcommit.
- Test disk p99 write latency, not just IOPS on a fresh volume.
A VDS is a perfectly reasonable buy when the provider is honest about placement. The mistake is comparing monthly prices between plans that are not selling the same thing. Price the usable core, measure steal time yourself, and the decision becomes arithmetic rather than marketing.




Leave a Reply
You must be logged in to post a comment.