Primetime ISP speeds: LocalNet vs Elisa vs DNA

Scheduled Ookla speed tests, 8 Feb – 24 Aug 2026 · 15,808 server measurements over 192 days with data · times in Europe/Helsinki · all rates in decimal Mbps

The evening dip, hour by hour

Where the congestion sits

Six months of primetime downloads, day by day

The second bottleneck: upload at 21:15

Month by month at the worst hour

Packet loss — the original complaint

The alert log tells the same story

In fairness to LocalNet: latency under load

New in the rotation: Bahnhof (Ylivieska/TPY)

A fourth connection — Bahnhof, subscribed 1000/1000 Mbps, at a second Ylivieska site — joined the schedule on 19 August. Its numbers should not yet be read as ISP performance: the site still has an old Cisco C897VA-M-K9 router, obtained for the 50/50 VDSL line the connection replaced, which tops out around 200–250 Mbps of routed traffic.

The first days of tests show a stable hardware-bound profile: ~215 Mbps toward the DNA Welho and Telia servers and ~300–350 Mbps toward Elisa and SuomiCom, identical at every hour, with no primetime dip, zero packet loss, and a well-behaved ~22 ms loaded latency.

Once the router is replaced with an EdgeRouter 4 (matching the other sites), Bahnhof joins the charts as a proper 1000/1000 comparison, and the swap date will be marked on the timeline.

Side-by-side summary

Timeline: complaint to (partial) resolution

From the support thread with LocalNet (“Pakettikatoa iltaisin”, Jun 2025 – Jun 2026) and from the measurement data.

Method

From 7 February 2026, a cron job has run the Ookla Speedtest CLI seven times daily — at 01:15, 07:15, 13:15, 17:15, 19:15, 21:15 and 23:15 (Europe/Helsinki); since 20 August two extra evening slots at 20:15 and 22:15 make it nine. Each run tests first the LocalNet connection, then Elisa (+1 min), then DNA (+2 min), and since 19 August also Bahnhof (+3 min).

The connections are at different sites — LocalNet in Kauhava (KAU), Elisa in Ylivieska (PRY), DNA in Espoo (ESP), Bahnhof at a second Ylivieska site (TPY) — behind identical customer equipment (Ubiquiti EdgeRouter 4, gigabit LAN) except Bahnhof (see its section). All test toward the same four fixed servers: SuomiCom (Espoo), DNA Welho, Telia and Elisa (all Helsinki). Differences in evening behavior therefore reflect the ISPs' networks, not the measurement setup; absolute idle latencies do differ with distance to Helsinki.

The data comes in two series: the original text logs (8 February – 12 August, one file per run) and, after a 6-day pause, a new JSON series (from 19 August, one Ookla JSON result per server via speedtest -f json -b -s <id>). This report covers 8 February – 24 August 2026: 15,808 measurements; 33 scheduled runs in the old series produced no result (measurement-host gaps, e.g. 18–20 July) and are excluded. The connection under test is identified from the ISP field reported by the test itself.

Subscribed rates: LocalNet 1000/1000, Elisa 1000/1000, DNA 600/100, Bahnhof 1000/1000 Mbps. Units: the old text logs report rates in mebibits per second (“Mibps”), while the JSON series reports bytes/s (decimal). All old-series rates have been converted ×1.048576, so every figure in this report is in decimal Mbps — the unit Ookla and the ISPs use. (The initial 19 August publication mistakenly read the old logs as decimal Mbps, understating all rates by ~4.9%; relative comparisons were unaffected.)

Packet loss isn't reported by every server/run (3,961 of 15,808 measurements lack it — most often toward the DNA Welho server); loss statistics use only tests that reported it. Latency figures are Ookla's idle and loaded (during transfer) measurements; for the JSON series the loaded figure is the IQM. All aggregates are medians unless noted, computed per measurement (server run).

The alert chart comes from a separate SmokePing mesh, independent of the speed tests; its master sits on a different line entirely — a second, separate Elisa connection at the Ylivieska PR site. Every 5 minutes it pings its targets from seven vantage points: Kauhava (behind the LocalNet line), Lappeenranta (on a DNA connection), Avaruus.net and Hetzner in Helsinki, AWS Stockholm, Linode London, and the master's own site. Targets include each site's router, the routers' uplink gateways, hosts in LocalNet's data center, and DNS infrastructure in Finland and abroad.

An edge-triggered Loss alert is raised when a target shows non-zero packet loss in two polling rounds within 30 minutes, and stays raised while the pattern keeps matching (pattern >0%,*4*,>0%, unchanged since June 2025; the target set is unchanged since 11 March 2026, apart from the dead data-center host described below, removed on 23 August).

The chart measures time in alert. Raised/cleared events are paired into episodes per target and vantage; each day's value is the number of 5-minute probe slots those episodes cover, summed over the paths that involve the LocalNet connection — measured from Kauhava, targeting the Kauhava routers or their uplinks, or targeting LocalNet's data center from any vantage point. Alerts still open when the SmokePing service restarts never log a “cleared”, so episodes are closed at service boundaries. A journal export is taken whenever the exporting script runs, which leaves the journal's first and last calendar days partial; the series covers complete local days only (15 Mar – 23 Aug here), so no bar stands for a fraction of a day.

Two exclusions keep the series about congestion. One data-center test host went silent on 5 April and has not returned; a dead host re-raises from every vantage after each service restart, which is an artifact, not congestion. And an episode is attributed to the far end when its target is simultaneously in alert from two or more non-Kauhava vantages — a 10-hour nameserver outage elsewhere on 26 June is the far end's problem even when the Kauhava path also alerts on it. Such episodes count under “all other paths”.

Raised-alert counts remain in the table view for reference. The source is the SmokePing master's systemd journal, whose retention horizon is 14 March 2026 — the worse February period predates it and survives only in the complaint thread's counts.

Changelog