Know what the numbers mean.

01 / Price and ordering

Annual prices are divided by 12 for display and still require full upfront payment. Prices and ordering use internal reference exchange rates; providers bill in the original currency. Taxes, setup fees, renewal and stock may be unknown. Recheck source prices older than 30 days.

02 / CPU and memory

Native CPU workloads: prime trial division up to 10000, 32×32 f64 matrix multiplication and SHA-256. Each has three independently warmed trials, 750 ms each by default; allocation and warmup are untimed. Per-trial rates, variation and cgroup throttling counters remain available. One thread then quota/affinity-aware multi-thread mode (max 8). Results retain verified answers, counts, durations and rates. A separate sustained prime load defaults to 20 seconds (max 120; --cpu-burn-s 0 disables), with actual interval rates and read-only CPU temperatures. Missing sensors remain unknown; rate changes alone do not identify thermal throttling. Independent Copy/Scale/Add/Triad kernels use fresh OS / visible-cgroup ancestor headroom to cap three arrays at 1/4 of capacity and half the available estimate, with an auxiliary reserve and explicit working set, one warmup and 10 timed iterations; traffic includes reads and writes. Three verified shuffled pointer-chain trials add access latency with an explicit working set; caches can affect it. Separate single-array sequential u64 reads/writes add ten timed, verified passes, counting logical bytes. These produce their own metrics, not Geekbench/sysbench or official STREAM scores.

03 / Storage

A 64 MiB anonymous file provides write+fsync and explicitly cached-read baselines. On supported Linux filesystems, random read, write and 50:50 mixed profiles run at 4K/64K/512K/1M, plus 1M sequential read/write: 14 direct I/O profiles plus four concurrent profiles with four synchronous QD1 workers, disjoint regions and a shared cap. Space checks and a 4 KiB write/sync/readback check are separate. Single-job QD1; each stops at 250 ms or 16 MiB, checked between operations. Another 28 Linux native AIO profiles add two-job QD64 mixed, random QD32, sequential QD8 and ten QD4 read/write block sizes from 512 B through 64 MiB. Each has a 256 MiB cap. V2 records a fresh budget for aligned data, typed latency arrays and auxiliary storage per profile; oversized fixed queues retain their parameters and are skipped. Queues and latency storage are allocated and first-touched before synchronized timing; compare separately from v1. Requested depth and per-job peak outstanding requests are separate; pending I/O drains after submissions stop and write rates include fsync. All 46 profiles preserve IOPS, throughput, P95 and failure reasons. Keep short datasets and sync policies distinct from fio/ATTO scores.

04 / Throughput and direction

Defaults: an 8 MiB Cloudflare nearest-edge download, a 4 MiB synthetic upload, IPv4/IPv6 TCP and available ICMP. Reports upload automatically only with explicit --submit consent. --network-targets accepts AFDATA targets:FILE or env:NAME, with at most 5 downloads, 2 uploads and 6 probes; --offline disables all network measurements. HTTP timings include DNS/TLS and are not iperf3. CDN edges have no fixed-region guarantee. --global-network adds native two-stream iperf3 at London, Singapore and New York on IPv4/IPv6, capped at 16 MiB or five seconds per direction. Uploads pair receiver bytes and receiver time. --network-diagnostics adds masked egress ASN, RFC5780 mapping/filtering/hairpin, RIS announced prefixes and AS neighbors, IX registrations and a 22-provider MX / SMTP / POP3 / IMAP matrix. NAT uses isolated ports and source checks; unsupported servers and unobserved hairpin stay unknown. Path positions do not establish commercial upstreams, and IX declarations do not measure VPS speed. Mail retains DNS, connection, protocol capability, certificate identity and fresh post-TLS evidence per endpoint, with up to 240 additional seconds and shared 8 MiB receive / 1 MiB send caps (DNS/MX metadata additional); no login or mail. Also measures cold IPv4/IPv6 connections to 40 websites and separate ICMP/TCP at 13 source-declared Telegram addresses. Three samples per pinned address retain DNS, TCP, TLS and HTTP header timings without redirects; 403/429/5xx do not establish unlocking. The service phase runs concurrently with mail and lasts at most 240 s; HTTPS TLS records are capped at 30 MiB receive / 3.75 MiB send total, excluding DNS and network-layer overhead. SpamCop/PSBL/DroneBL DNS controls and an official Tor snapshot retain unknown, stale, missing-key and unsupported-family states. Optional owner IPapi.is credentials add attributed vendor flags, never a composite risk score. Also checks 60 anonymous media/AI and application services: region flags, frontend and catalog/manifest metadata remain distinct from playback/entitlements. No login, playback or generation; up to 96 MiB consumed bodies over a 240-second phase, concurrent with other diagnostics. Another 138 indexed functions remain pending. --china-network also runs a 31-province three-carrier IPv4/IPv6 TCP matrix: ten rounds or 44 with --china-deep, retaining every sample, start offset, DNS candidate and endpoint selection; it adds public-carrier HTTPS transfers with remaining-budget legacy TCP fallback (96 MiB down / 48 MiB up total), 15 ICMP targets, small/large TCP packets at 18 targets and six Linux UDP return routes. Existing NET_RAW permits 12 additional TCP/ICMP paths with three samples per TTL and bounded RIS origin-ASN lookups. Missing raw permission is explicit; missing router replies are not application packet loss and origin ASNs do not certify routing products. Without NET_RAW, TCP-connect fallback is explicit. HTTP pre-read bytes stay charged and successful directions are not repeated. Attempts and negative controls remain visible; HTTPS upload metering is distinct from server byte acknowledgement. Transfer directions describe data flow; mainland-initiated probe evidence is separate. Carrier and region labels retain their sources; short samples do not establish nationwide coverage or port capacity.

05 / Evidence, freshness and privacy

Seed plans contain provider claims, no invented benchmarks. Reports include runner version, actual server region and UTC timestamp; after 30 days they become stale. Reports omit SSH target, hostname, full public IP and credentials; optional diagnostics retain /24 or /48 prefixes only. Imported records are public, so the operator can review them first. JSON and Markdown preserve unknown values and explicit statuses.

06 / Editions and reports

English and Chinese catalogs and editorial copy are maintained separately; Chinese pages focus on the three mainland carriers. Empty filters remain empty. Measured means a report exists, not that a provider passed certification.

07 / Rebuildable single-run scores

Performance and China network use unbounded indices with reference 100. Performance retains its fixed weighted geometric mean. Network contributions add with fixed weights: latency 40%, connection success 20%, download and upload 20% each. Any valid measurement produces an observed contribution index; missing values stay unknown without redistributing weights. Partial indices are labeled observed lower bounds with coverage counts. IPv4/IPv6 each carry 50% in carrier indices, followed by fixed equal weights for three carriers in the China region index. Readings retain their family and initiation directions stay separate. RST is not acceptance. Algorithm upgrades preserve source reports.

Read the algorithm, calibration and component weights ↗

Upstream test coverage

53 test groups from YABS, GoECS and NetQuality are reviewed with sources and parameter differences. The unlock index has 198 exported entries including aliases/fallbacks, 60 partially covered and 138 unimplemented. Two pinned GoECS dependency sources returned 404 and remain unaudited. Native Linux asynchronous disk tests, repeated 31-province measurements, RFC5780 NAT stage evidence and optional read-only hardware inventory are implemented. Proprietary hardware health, route branding, commercial upstream relationships and media/AI unlocking remain incomplete.

VPS buying and report questions