Benchmarks: measured on whole nodes

This page shows load-test results for cdnca nodes with 2, 4 and 8 cores, so you can size a node and plan capacity from your traffic.

Summary

Scenario: reverse proxy plus access control worst path (5 rules × 5 conditions, hit on the last condition of the 5th rule). HTTPS uses an RSA 2048 certificate.

Node specHTTP sustainable throughput (QPS)HTTPS sustainable throughput (QPS)
2 cores (Intel Xeon E5-2686 v4 @ 2.30GHz)2,0002,000
4 cores (Intel Xeon E5-2686 v4 @ 2.30GHz)4,0004,000
8 cores (Intel Xeon E5-2686 v4 @ 2.30GHz)8,0008,000

Detailed results

Same scenario. Each spec lists peak and sustainable figures for HTTP and HTTPS.

Node specHTTPHTTPS
2 cores (Intel Xeon E5-2686 v4 @ 2.30GHz)
Peak throughput:2,248 QPS
P99 latency 10,475.36 ms
Error rate 0.00%
Sustainable throughput:2,000 QPS
P99 latency 21.75 ms
Error rate 0.00%
Peak throughput:2,129 QPS
P99 latency 458.33 ms
Error rate 0.00%
Sustainable throughput:2,000 QPS
P99 latency 21.68 ms
Error rate 0.00%
4 cores (Intel Xeon E5-2686 v4 @ 2.30GHz)
Peak throughput:4,327 QPS
P99 latency 414.36 ms
Error rate 0.00%
Sustainable throughput:4,000 QPS
P99 latency 26.44 ms
Error rate 0.00%
Peak throughput:4,207 QPS
P99 latency 447.02 ms
Error rate 0.00%
Sustainable throughput:4,000 QPS
P99 latency 42.76 ms
Error rate 0.00%
8 cores (Intel Xeon E5-2686 v4 @ 2.30GHz)
Peak throughput:8,396 QPS
P99 latency 220.99 ms
Error rate 0.00%
Sustainable throughput:8,000 QPS
P99 latency 35.99 ms
Error rate 0.00%
Peak throughput:8,342 QPS
P99 latency 230.05 ms
Error rate 0.00%
Sustainable throughput:8,000 QPS
P99 latency 27.28 ms
Error rate 0.00%

Protection scenarios

Sustainable / peak QPS at 2, 4 and 8 cores for each protection mix (each cell shows “sustainable / peak”, in QPS).

Configuration2 cores4 cores8 cores
Reverse proxy (origin fetch)2,000 / 2,4764,500 / 4,7569,000 / 9,317
Cache hit2,000 / 2,6584,500 / 4,91210,000 / 9,996
CC, 1 rule2,000 / 2,3254,500 / 4,5948,000 / 9,088
CC, 10 rules all counting2,000 / 2,1564,000 / 4,2137,000 / 8,292
CC worst path (10 rules × 3 conditions, 2-key count)2,000 / 2,3184,000 / 4,5138,000 / 8,786
Access control worst path (5 rules × 5 conditions)2,000 / 2,2484,000 / 4,3278,000 / 8,396
Managed WAF standard (CRS PL1)400 / 652500 / 1,2471,000 / 2,438
Managed WAF strict (CRS PL2)200 / 400250 / 760— / 1,474
Managed WAF paranoid (CRS PL3)200 / 368250 / 709— / 1,389
JS challenge (every request gets the challenge page)3,500 / 4,000≥ 6,000 / ≥ 6,000≥ 12,000 / ≥ 12,000
Automatic defense on (idle state)2,000 / 2,4084,000 / 4,6188,000 / 9,218
Everything on (ACL 5×5 + CC worst path + CRS PL1 + automatic defense)400 / 615750 / 1,1921,000 / 2,305
HTTPS reverse proxy (ECDSA P-256)2,000 / 2,1824,000 / 4,1978,000 / 8,473
HTTPS reverse proxy (RSA 2048)2,000 / 2,1294,000 / 4,2078,000 / 8,342
HTTPS, everything on400 / 588500 / 1,1501,000 / 2,117

“—”: a dash in the sustainable position means even the lowest step was outside the sustainable criteria; a dash for the whole cell means it was not measured this round. “≥”: the top step was still sustainable, so the real figure is higher.

100,000 sites

The node agent’s dynamic routing runtime loading 10,000 and 100,000 sites (2 domains each) with the OpenResty from the same 1.0.35 release, 2 workers, a 10-core 6 GB container.

Item10,000 sites100,000 sites
First load of all sites2273.1ms18062.2ms, 17958.7ms
Change 1 site until live812.4ms1436.2ms, 1572.6ms
Reload on site changeNoNo, No
Warm request p50 (1 client)707us820us, 826us
Warm request p99 (1 client)1285us1781us, 1628us
Warm request p99 (16 clients)5882us12563us, 12428us
OpenResty total PSS (after load)268.0MiB379.5MiB, 382.6MiB
Node restart until serving990.6ms9202.1ms, 8725.9ms

Test environment and topology

  • Host CPU: Intel Xeon E5-2686 v4 @ 2.30GHz (KVM, one thread per core), 32 vCPU / 31 GB in total.
  • Node, origin and load generator run on the same host, split by cpuset: node 2 / 4 / 8 cores, origin 4 cores, load generator (k6) 16 cores, over loopback.
  • Software: cdnca 1.0.35 (source commit 6d69d0480), OpenResty 1.31.1.1, Ubuntu 24.04 in the container; worker count equals the node core count (worker_processes auto).
  • Load tool: k6.
  • Configuration: sites and rules are created on a control plane, published through the normal pipeline to a real node, and the published site objects, index, nginx.conf, certificates and rule plans are copied from that node and replayed.

Topology: k6 (client) → node → origin (keepalive)

Metric definitions

  • Target QPS: the request arrival rate configured in k6 constant-arrival-rate.
  • Achieved QPS: the rate at which requests actually complete.
  • Peak: the highest achieved QPS over all steps.
  • Sustainable: the highest target rate where error rate < 0.1%, P99 ≤ 50 ms and achieved ≥ 99% of target; use it for capacity planning.
  • Saturation: the first step where raising the target no longer raises achieved QPS in step and P99 rises above twice the sustainable P99.
  • Error rate: 5xx, timeouts and connection errors only; 4xx and challenge pages returned by protection by design are not counted.

Method

  • Request: GET a 1 KB static page, Accept-Encoding: identity, desktop Chrome User-Agent; the origin answers no-store (origin fetch) or no cache headers (cache-hit row).
  • HTTP: HTTP/1.1 keepalive. HTTPS: TLS 1.3 + HTTP/2 with connection reuse. Origin: HTTP/1.1 keepalive. CC and automatic-defense rows rotate 4,096 X-Forwarded-For addresses.
  • Stepped load: k6 constant-arrival-rate, 30 s per step with 5 s gaps; the first 5 s of each step are dropped from CPU figures. Step targets follow the core count: 2 cores 500 to 4,000 in steps of 500; 4 cores 1,000, 2,000, 3,000, 3,500, 4,000, 4,500, 5,000, 6,000; 8 cores 2,000, 4,000, 6,000, 7,000, 8,000, 9,000, 10,000, 12,000.