Wings vs Manta benchmarks
Fresh current-release measurements of daemon APIs, file operations, game lifecycle and resource use.
Measured 2026-11-22: Manta v1.3-beta versus Wings v1.13.5, compiled from unmodified release tags with Go 1.27.4 and matching build flags. Both use the same runtime image and resource limits. The homepage shows three fixed metrics; all measured workloads, including regressions, appear here.
Raw results · Exact setup and runner files · Repository harness
Sources and environment
| Setting | Wings | Manta |
|---|---|---|
| Release | v1.13.5 | v1.3-beta |
| Exact source | 3236480d078c09b823b8a8f21dab76c4029608a5 | 25dac4d729af64f53f711bf3a3f974b3c6a7b927 |
| Go compiler | go1.27.4 | go1.27.4 |
| Daemon CPU quota | 2 cores | 2 cores |
| Daemon RAM limit | 1024 MiB | 1024 MiB |
| GOMAXPROCS | 2 | 2 |
| Native Panel backend | pterodactyl_v1 v1.15.2 | aquadactyl v1.3-beta |
Unmodified tagged source, identical Go 1.27.4 compiler/build flags and common Alpine runtime image. Each daemon has a two-core quota, 1 GiB RAM, GOMAXPROCS=2. Games have matching pinned images and configured limits. Private nested Docker engine is limited to 16 cores/12 GiB and publishes no host ports. Manta uses Aquadactyl; stock Wings uses its native Pterodactyl v1 backend (Wings page=0 bootstrap is incompatible with Aquadactyl validation). No production-code compatibility patch is applied. Panel and daemon workloads do not overlap.
Both executables are rebuilt from released source, rather than comparing a custom Manta binary to an upstream prebuilt Wings image. Binary and Go dependency hashes, runtime image IDs and effective game-container limits are recorded. The native backend pairing avoids changing stock Wings' zero-based bootstrap pagination. HTTP/file timings call the daemon directly; SFTP authentication and bootstrap are not part of transfer timers.
The host is Intel(R) Core(TM) i7-14700KF, 28 logical CPUs and 64 GiB RAM. The private engine is Docker 29.8.1; the load generator has four CPU cores and 1 GiB RAM. Existing unrelated workstation containers remain running.
Matching game fixtures
Each daemon manages one Paper 26.3 build 257 server and one Bedrock 1.26.66.1 server. The same downloaded bytes and image digests are used on both sides. Paper has 1536 MiB configured memory and Bedrock 512 MiB; both have a 200% CPU limit. Effective cgroup memory includes the daemons' normal overhead allowance and is checked for parity. Both use the same seed and properties apart from assigned ports. No authenticated players or bots participated.
Four start/stop rounds per game and daemon in alternating order. Fresh worlds are initialized by the first start, which is retained as a discarded warm-up; only the next three warm starts contribute to the published lifecycle medians. Timers cover power request through daemon-reported state, with 100 ms polling; shutdown is not a separately timed OS process exit. Identical cached binaries/image layers, seed and properties; normal configuration sync and registry digest checks remain in the timed power path. Bedrock explicitly uses transport=raknet for the UDP query workload.
API throughput and latency
Three alternating repetitions per daemon, endpoint and concurrency (1/16/32). Five-second closed-loop authenticated HTTP/1.1 windows with persistent connections; one discarded request before each window. No new requests after the deadline; in-flight requests are drained. Throughput is successful requests per actual elapsed second. Body receipt is included, percentile sorting and publication are excluded. Timed responses must be HTTP 200. This replaces the old 400-request batches; do not claim a controlled before/after comparison across methods.
Throughput is the median of three per-window rates. The p50/p95 columns are medians of the per-window percentiles, not pooled percentiles. The measured runner uses the lower middle order statistic for p50 and floor((n−1)×0.95) for p95. All 4,654,944 measured API requests completed with zero errors. The directory size and running server state are checked outside each timed window.
| Endpoint / concurrency | Wings req/s | Manta req/s | Wings p50 ms | Manta p50 ms | Wings p95 ms | Manta p95 ms |
|---|---|---|---|---|---|---|
| Server status / 1 | 6936.844 | 7362.904 | 0.138 | 0.132 | 0.198 | 0.184 |
| Server status / 16 | 65224.959 | 82869.682 | 0.183 | 0.147 | 0.720 | 0.544 |
| Server status / 32 | 64412.451 | 82350.267 | 0.369 | 0.290 | 1.480 | 1.153 |
| List 1000 files / 1 | 133.258 | 140.087 | 7.466 | 7.096 | 8.327 | 7.841 |
| List 1000 files / 16 | 230.273 | 257.926 | 60.876 | 52.094 | 146.445 | 144.258 |
| List 1000 files / 32 | 230.367 | 259.663 | 116.914 | 100.375 | 326.807 | 289.994 |
Startup and shutdown
Three measured warm repetitions after excluding the first start. Values are medians in milliseconds; 100 ms polling limits precision.
| Measurement | Wings ms | Manta ms |
|---|---|---|
| paper: warm startup | 6650.361 | 6550.823 |
| paper: reported offline | 806.719 | 705.830 |
| bedrock: warm startup | 1108.450 | 1108.433 |
| bedrock: reported offline | 907.008 | 806.595 |
File and event operations
16 MiB shared random payload, three repetitions with five samples each for HTTP and SFTP read/write. SFTP authentication is outside transfer timers and uses each daemon's native Panel backend. Hash-verified reads and archives; 1000 x 256-byte directory fixture; 100 console lines delivered to 16 subscribers; automatic stats validated for 16 viewers per game/daemon.
Values are medians of per-run measurements (milliseconds). Integrity checks occur outside transfer timers.
| Measurement | Wings ms | Manta ms |
|---|---|---|
| http write 16m | 10.849 | 10.196 |
| http read 16m | 20.427 | 9.725 |
| sftp write 16m | 78.731 | 76.661 |
| sftp read 16m | 40.111 | 39.117 |
| chmod 1000 | 4.101 | 3.984 |
| rename 1000 | 13.558 | 11.805 |
| Compress 16 MiB + 1000 files | 15.064 | 17.690 |
| Extract archive | 45.225 | 43.193 |
| Deliver 100 log lines to 16 subscribers | 3.771 | 3.177 |
All 64 automatic-stat viewer checks passed: 16 viewers for each of four game servers, with three unsolicited samples each and advancing uptime. This checks the observed stream, not unlimited concurrency or production reliability.
Memory and command load
Fifteen Docker working-set samples at two-second intervals for each of six containers, once idle and once under matching game command load. Mean is arithmetic mean; raw sample values and observed peaks are retained. Working set subtracts inactive file cache, not RSS. Commands alternate 4096 blocks between stone and air, five iterations/second per node across both games, with matching forced/ticking areas. This is not real-player load.
Daemon memory excludes game containers, Panel, database, Redis and host overhead. No forced daemon GC occurs between workloads. The command workload is synthetic and deliberately small; these figures are not a player-capacity benchmark.
| Mean working set (MiB) | Wings | Manta |
|---|---|---|
| idle: daemon | 51.753 | 50.604 |
| idle: paper | 739.692 | 738.263 |
| idle: bedrock | 189.813 | 189.797 |
| command load: daemon | 43.877 | 47.817 |
| command load: paper | 763.033 | 764.760 |
| command load: bedrock | 257.992 | 257.751 |
| Local game query p50 (ms) | Wings | Manta |
|---|---|---|
| idle: paper | 1.439 | 0.967 |
| idle: bedrock | 0.346 | 0.301 |
| command load: paper | 0.859 | 0.830 |
| command load: bedrock | 0.327 | 0.295 |
Each game-query row contains 50 sequential local Java status or Bedrock UDP queries. These are server-query round trips, not gameplay latency. CPU core-seconds, sample arrays, observed peaks and Paper tick output are retained in the raw report.
Reproduction and limitations
The website repository contains the exact guarded setup, provisioning and workload sources. Follow its README with clean tagged checkouts and fresh disposable databases. Game binaries are downloaded from their publishers and are not redistributed with the report. Private account passwords, node tokens and configuration files are excluded from public artifacts.
Single shared Windows/WSL2 workstation with unrelated existing containers, no CPU pinning or repeated clean-machine trials. Linux container networking; no TLS/WAN/mobile/endurance test or authenticated players. Daemon and game memory are separate; this is a controlled local distribution comparison, not general hosting capacity. No new security audit, race test or full coverage run is claimed by this performance rerun.
November 2 results, its historical Manta source patch, coverage report and microbenchmarks remain available. Those security/coverage/microbenchmark runs were not repeated as part of this current-release performance comparison. Older 400-request API batches and existing worlds differ from this rerun; do not present differences across dates as a controlled optimization experiment.
v1.3-beta directory-listing follow-up
Manta v1.3-beta includes a separate component optimization measured on 22 November 2026 against revision 80f84fbda9a7513c995d5a0456085df5adcd9c71. Metadata is collected relative to the open directory descriptor, avoiding repeated parent-path traversal while retaining root-confined, nonblocking MIME inspection.
| Component case | Before | After | Time reduction |
|---|---|---|---|
| 1,000 regular files, root read | 5.250 ms | 4.395 ms | 16.3% |
| 1,000 regular files, nested read | 9.355 ms | 6.710 ms | 28.3% |
| Nested listing plus API JSON encoding | 10.068 ms | 7.358 ms | 26.9% |
The root read used 45.4% fewer allocations, from 22,021 to 12,020 per operation. These are medians of five alternating 500 ms batches per variant after a discarded warm-up, using Go 1.27.4, the same fixture/dependencies, 2 CPUs, 2 GiB memory and GOMAXPROCS=2 on one Windows/WSL2 workstation. Nested fixtures and distinct file contents are included. The test used warm Linux temporary storage, without competing test/build jobs; unrelated workstation services remained running.
This does not replace the live HTTP/Wings comparison above or establish game/player capacity. The next full benchmark must measure the released daemon under the same complete workloads. See the method and all eight cases and raw samples.