Benchmarks

API readslist, detail, stats, nodes · ↑ requests/s

GoWave1,169

Remnawave555

SubscriptionsXray JSON, Mihomo, sing-box · ↑ requests/s

GoWave561

Remnawave400

Memory under loadpanel container, median · ↓ MiB

GoWave60

Remnawave1,036

Startupuntil the API answers · ↓ ms

GoWave209

Remnawave7,061

GitHub Actions (4 vCPU), Docker, 2 CPU and 2 GiB per panel, PostgreSQL 17 (1 CPU) each, 1,000 users, 32 concurrent clients. Median of 3 20 s runs. Remnawave 3.4.5.

Setup

• Remnawave 3.4.5 (remnawave/backend:3.4.5) versus GoWave at commit 349f8ed.
• Run on GitHub Actions on a runner with 4 vCPU (AMD EPYC 7763) and 16 GB RAM. Logs, raw results and the workflow are public, so the run can be repeated.
• Docker, each panel limited to 2 CPU and 2 GiB, each with its own PostgreSQL 17 on 1 CPU and identical settings. Remnawave also gets Valkey (0.5 CPU), which it requires.
• The same data, created through each panel’s API: 1,000 users, 20 hosts, a profile and a squad.
• Load is generated by k6 1.6.1: each iteration is one HTTP request. Responses are validated: an error, an empty subscription or HTML counts as a failure.
• Run order alternates and is shuffled; every run has a warmup. Each run lasts 20 seconds, three repetitions, tables show the median.
• Before every run both panels’ databases get CHECKPOINT and VACUUM ANALYZE, so one run does not slow down the next.
• Latency is server response time (k6 timings.duration); p99 is the median across runs.

Throughput

Concurrency mode: a fixed number of clients send requests back to back, each one right after the previous response. This shows how many requests per second a panel actually sustains.

Workload	Clients	GoWave, requests/s	p99, ms	Remnawave, requests/s	p99, ms	GoWave / Remnawave
Reads (mix)	1	472	4.0	282	7.1	1.67×
Reads (mix)	8	1,084	15.8	573	38.4	1.89×
Reads (mix)	32	1,169	62.6	555	122.0	2.11×
Reads (mix)	128	1,181	254.1	553 †	556.8	2.14×
Subscriptions	1	270	4.6	225	4.8	1.20×
Subscriptions	8	557	25.6	370	33.4	1.51×
Subscriptions	32	561	109.5	400	131.1	1.40×
Subscriptions	128	555	451.6	395 †	579.4	1.40×
Mixed	1	360	4.1	251	7.0	1.44×
Mixed	8	779	20.5	463	32.4	1.68×
Mixed	32	833	80.2	470	134.3	1.77×
Mixed	128	823	359.1	475 †	540.9	1.73×
Writes	1	495	2.7	285	5.0	1.74×
Writes	8	1,449	11.2	634	19.8	2.28×
Writes	32	1,467	51.8	652	76.5	2.25×
Writes	128	1,450	225.7	670	288.9	2.16×
User list	1	242	4.5	157	8.6	1.54×
User list	8	507	28.0	227	76.9	2.23×
User list	32	508	119.1	213	324.6	2.39×
User list	128	500	513.4	211 †	1,111.8	2.38×
User detail	1	827	1.7	564	2.6	1.46×
User detail	8	2,242	7.8	1,170	10.9	1.92×
User detail	32	2,351	32.2	1,206	40.6	1.95×
User detail	128	2,333	125.2	1,193	165.3	1.95×
Dashboard stats	1	663	2.1	287	5.0	2.31×
Dashboard stats	8	3,046	5.5	637	22.3	4.78×
Dashboard stats	32	4,155	21.1	650	78.4	6.39×
Dashboard stats	128	4,303	70.7	649	294.1	6.63×

† marks runs where Remnawave failed the SLO: p99 under 500 ms and under 1% errors.

Latency at a fixed load

Rate mode: exactly 100 or 500 requests per second. Both panels keep up, so the comparison is about response times.

Workload	Requests/s	GoWave p50 / p99, ms	Remnawave p50 / p99, ms
Reads (mix)	100	1.4 / 4.3	3.2 / 8.1
Reads (mix)	500	1.5 / 10.4	15.8 / 487.4
Subscriptions	100	2.4 / 5.1	2.9 / 7.2
Subscriptions	500	8.3 / 131.8	584.3 / 10,000.7

Memory and startup

• Memory (median across all runs, whole panel container including workers): GoWave about 60 MiB, Remnawave about 1,036 MiB.
• Startup until the first authenticated API response (median of 10 trials): GoWave 209 ms, Remnawave 7.1 s.

Worth knowing

• Dashboard stats at 32 clients and more (up to 6.4×) comes from coalescing concurrent requests: while one computation runs in the database, the others wait for its result. Such a stampede on one endpoint is rare; the everyday difference shows with a single client: 2.3×.
• Writes from 32 clients up are limited by GoWave’s PostgreSQL, not by the application: the database, capped at 1 CPU, is throttled in 81-85% of periods, the panel container is not.
• Nodes are not part of the test. Node management, traffic collection, HWID and real VPN clients were not measured, only the panel API.
• This is a shared GitHub Actions virtual machine, not a capacity study: the active panel, its database and k6 share 4 cores, and runs are short. The numbers show the ratio, not absolute limits.

All charts

Where the numbers come from

The tests live in the public gowave-bench repository: k6 scenarios, both panels in Docker, data preparation and charts. Runs happen on GitHub Actions: the workflow builds GoWave from source, pulls the Remnawave image and compares them on the same virtual machine.

Every run’s logs, machine details and raw results (in the artifacts) are public. The numbers on this page come from run 40386568113. To check them, fork the repository and run the Benchmark workflow with the full profile.