ARGUS EXTERNAL HTTPS BENCHMARK — 27 OCTOBER 2026 (America/Sao_Paulo)

Purpose: compare complete successful responses within a fixed 500-ms latency budget, rather than request volume alone.
36 measured trials across three configurations: 99,000 scheduled requests; 97,142 sent. A separate 1,000-request exploratory run is included in the raw archive and excluded from those totals. Warmup requests are excluded from measurement totals.

Environment: one 1-vCPU / nominal 512-MB Linux cloud server (approximately 458 MiB usable RAM), Ubuntu 24.04 and NGINX. An external Windows workstation generated requests directly to the origin with correct domain Host/SNI and TLS certificate validation. Cloudflare was bypassed. The full 33,753-byte static homepage was read and checked against SHA-256 5be0fe1055c583cf76d3921ec115d7dbf6380111bdbc4b7f8d782130dcde3494. Compression was disabled using Accept-Encoding: identity.

Method: fixed scheduled arrivals at 50, 100 or 200 requests/second, reusable HTTP/1.1 connections, a 2-second timeout, 3 seconds of warmup, then 20 seconds of measurement. Three trials per configuration/rate. Profile order rotates reference/tuned/argus, tuned/argus/reference, argus/reference/tuned. The initial suite used a 64-active-request client cap. A separate nine-trial 200-RPS validation used a 128 cap after initial high-load trials encountered the client's ceiling. Both sets are retained.

Latency runs from scheduled arrival until the complete body is received. Percentiles use nearest rank across valid HTTP 200 responses; timeouts are reported separately. Requests dispatched over 50 ms late or prevented by the active-request cap are recorded as generator misses, not server timeouts. Useful RPS is valid responses completed within 500 ms divided by the scheduled measurement duration. Percent within budget uses all scheduled requests as its denominator. Timeout rate uses sent requests as its denominator.

Configurations:
- Configured reference: worker_connections 768; worker_rlimit_nofile override removed; explicit multi_accept/epoll removed; somaxconn 4096; TCP cubic. This is not a fresh default OS image; other existing kernel settings remain constant.
- Tuning only: captured production settings, including worker_connections 16384, worker_rlimit_nofile 65535, multi_accept/epoll, somaxconn 65535 and BBR.
- Tuning + Argus telemetry: the same tuning, plus the actual telemetry script in a dedicated temporary service monitoring HTTPS port 443; per-source connection threshold 200; ban TTL 30 seconds; purge on exit.
The existing XDP program, with an initially empty blocklist, remained attached in every configuration. These trials do not isolate XDP's performance benefit. No attack traffic was generated.

Results are pooled across three 20-second runs per row. Raw per-request data and per-run summaries are supplied. Times in those files are UTC; the test date above is local to America/Sao_Paulo.


MAIN COMPARISON

Profile | Offered RPS | Within 500ms | Useful RPS | p95 ms | p99 ms | Timeouts / sent | Generator misses
Configured reference | 50 | 99.97% | 49.98 | 319.9 | 344.5 | 0 / 3000 | 0
Tuning only | 50 | 98.27% | 49.13 | 242.4 | 701.4 | 0 / 3000 | 0
Tuning + Argus telemetry | 50 | 99.83% | 49.92 | 236.8 | 279.3 | 0 / 3000 | 0
Configured reference | 100 | 99.07% | 99.07 | 334.3 | 493.2 | 0 / 6000 | 0
Tuning only | 100 | 99.27% | 99.27 | 294.9 | 415.9 | 0 / 6000 | 0
Tuning + Argus telemetry | 100 | 100.00% | 100.00 | 246.6 | 312.1 | 0 / 6000 | 0
Configured reference | 200 | 99.67% | 199.33 | 393.7 | 457.8 | 0 / 12000 | 0
Tuning only | 200 | 95.41% | 190.82 | 447.6 | 873.0 | 5 / 11877 | 123
Tuning + Argus telemetry | 200 | 95.32% | 190.63 | 402.6 | 1367.1 | 8 / 11768 | 232

INTERPRETATION
At 100 RPS, tuning plus telemetry completed 6,000/6,000 responses within 500 ms, with no timeouts or generator misses. The configured reference completed 99.07% within budget. Pooled p99 was 312.1 ms versus 493.2 ms. This is an observed workload-specific latency result.

At 200 RPS with the 128-cap client, the configured reference completed 99.67% of scheduled requests within budget; tuning plus telemetry completed 95.32%, with eight timeouts and 232 generator misses. These trials do not establish an Argus capacity advantage at 200 RPS. Do not interpret client misses as server rejection counts, or successful-response percentiles without the timeout and miss columns.

One host, one external internet path, short trials and one static page limit generalization. Profile rotation reduces timing bias but does not establish causality or remove internet/host variability. The comparison does not prove operating-cost savings, a maximum sustained capacity, an SLA, DDoS resistance or that telemetry caused the observed latency improvement. This is a tested evaluation configuration, not certification of the packaged installer, persistent services or reboot behavior.

INITIAL 200-RPS TRIALS (64-CAP CLIENT; RETAINED FOR COMPLETENESS)

Configured reference: 99.57% within 500 ms; 199.13 useful RPS; p95 224.8 ms; p99 372.8 ms; 0 timeouts; 23 generator misses.
Tuning only: 90.69% within 500 ms; 181.38 useful RPS; p95 442.2 ms; p99 647.3 ms; 2 timeouts; 813 generator misses.
Tuning + Argus telemetry: 92.16% within 500 ms; 184.32 useful RPS; p95 397.5 ms; p99 575.1 ms; 0 timeouts; 667 generator misses.

Earlier local-loopback and flood demonstrations are separate experiments. They are not included in the 36-trial external suite.
Contact: william@prometheusdynamics.ca
