Repeatable tests.
Inspectable evidence.
A benchmark should help you understand a workload, a machine, and the conditions behind a result.
Start with the deployment.
A specific instance is the unit of evidence.
For a controlled plan summary, provision three fresh instances independently across different days. Observe each for approximately 24 hours and repeat workloads in defined windows. Fresh instances do not guarantee different physical hosts.
Record the provider, plan, region, CPU, RAM, storage, OS, kernel, virtualization, agent release, and workload configuration. Unknown fields stay unknown. Operator-supplied values remain distinct from detected facts.
Every workload. Together.
| Area | Workload | What the report shows |
|---|---|---|
| CPU | Single-thread and all-vCPU sysbench; compression | Throughput, runtime, variability |
| Memory | Fixed working set exceeding CPU cache | Read, write, and copy bandwidth |
| Disk | fio direct I/O: separate read, write, and mixed jobs | MB/s, IOPS, p50/p95/p99 latency |
| Network | Controlled iperf3 endpoints; latency probes | Direction, source, destination, throughput, loss |
| Database | PostgreSQL pgbench; in-memory and storage-bound | Transactions/sec, latency, errors, durability |
| Consistency | Repeated measurements across the test window | Median, range, degradation, CPU steal |
Run workloads sequentially so they do not compete with one another. Pin tool versions and publish exact job settings. Storage tests use isolated files, never arbitrary raw devices. Database tests keep durability enabled.
Compare like with like.
Keep shared, dedicated, burstable, and ARM instances in separate cohorts. Different CPU and memory allocations are visibly labelled. A regional comparison does not imply identical physical locations.
Publish raw measurements and workload-specific comparisons. Missing data is missing, never zero. There is no opaque overall score.
Real plan prices must include native currency, regular monthly cost, mandatory extras, transfer terms, tax treatment, and verification date. Test bills and retail monthly pricing are separate facts. Current sample prices are illustrative.
A person approves publication.
The agent saves the full local report and submits to private intake. Automated validation checks structure, checksums, versions, completeness, and duplicates. Passing those checks does not grant approval.
A reviewer inspects every category, detection evidence, failures, and artifacts on a continuous report. Publication binds to that exact approved revision. Any changed content needs another review. Only sanitized, approved snapshots and artifacts can become public.
Explore the benchmark agentRead the boundaries.
Three deployments are a small sample. They cannot establish provider-wide performance, long-term uptime, or the experience of every customer. Twenty-four hours describes short-term consistency only.
Network values apply to the named endpoints and directions. Tool settings, datasets, operating systems, time of day, and hardware placement can affect results. Price checks and benchmark dates are recorded independently.
External submissions retain their source classification. Client checksums establish integrity, not the truth of provider identity or execution. Manual review is editorial scrutiny, not hardware attestation.
Make the evidence visible.
Provider marketing specifications and observed measurements are separate. Valid slow runs remain in the evidence. Corrections need a visible reason and a new reviewed revision; historical reports are not silently overwritten.
Any future commercial relationships or referral links must be disclosed alongside the relevant recommendations. A provider’s commercial participation must not determine test results.
Read the publication changelog ↗