EdgeRush

← Home

Performance

EdgeRush, NGINX and HAProxy on like for like hardware, with the same workload. Results from 4 October 2026.

Requests per second

Higher is better. PGO is a separate, profile-guided EdgeRush build. Scroll the table horizontally on smaller screens.

Median throughput; k = 1,000 requests/s. 4 × 100 = four connections, up to 100 streams each.
ScenarioEdgeRushEdgeRush PGONGINXHAProxy
HTTP/1 over TLS38.4k46.7k37.8k36.7k
HTTP/2 over TLS, 4 × 10061.6k63.1k45.3k41.0k
HTTP/1, 256 connections48.2k60.4k52.9k47.5k
HTTP/2, 4 × 10063.1k62.7k44.6k48.4k
HTTP/2, one hot connection30.4k39.1k21.9k21.5k
gRPC, one hot connection25.1k26.7k14.0k15.1k
gRPC over TLS, saturation30.1k30.0k20.0k25.2k
HTTP/3, 4 × 10051.5k58.6k59.3k55.6k
HTTP/3, 256 connections33.8k36.5k30.8k35.3k
HTTP/3, one hot connection25.2k32.2k27.0k21.9k

NGINX's gRPC-over-TLS run included 980 non-2xx responses. HTTP/3 with four connections is sensitive to uneven worker placement. Small differences should be treated cautiously: the summary reports medians without repetition ranges.

Methodology

Hardware: Intel Core i7-10610U, 4 cores / 8 threads; 32 GB DDR4 at 2667 MT/s. CPU capped at 1.8 GHz, performance governor, turbo disabled. Ubuntu 26.04 LTS, Linux 7.0.0-28.

All proxies ran on the same machine: four workers on two physical cores, with one core each for the generator and backend, over loopback. Each scenario ran for 10 seconds, repeated three times in interleaved order. Routing, header handling and upstream keep-alive were configured for equivalent work; access logging was off.

EdgeRush revision ff377cd; NGINX 1.28.3; HAProxy 3.2.9. These results predate access logs and measure small-response throughput, not production capacity. Read the harness and configs.