HTTP (L7) proxy benchmark · open-loop ramp

request throughput 20260813-190516

Every proxy driven through the identical linear ramp (200→100000 req/s across 1,000 connections, zrk, open-loop, coordinated-omission corrected) as an HTTP/1.1 reverse proxy. Throughput & CPU span the full ramp. The table's p50 / p99 are each proxy's latency at a common 2,000 req/s reference load — a shared, light, sub-knee rate where the number is per-request cost, not queueing. The full latency-vs-load story is the p99 pane and the distributions below, each line stopping where that proxy stops keeping up.

proxy	max sustained req/s	p50 @ 2k · ms	p99 @ 2k · ms	peak mem · MiB
zoxy 0.2.1	42.1k	0.4	1.9	—
nginx 1.31	18.2k	0.5	1.1	—
haproxy 3.4.3	17.1k	0.4	2.8	—
pingora 0.8	12.8k	0.5	6.1	57.1
envoy 1.33	8.1k	0.6	3.3	78.6

"peak mem" is cAdvisor's sampled container RSS for the run — kernel, allocator, and page-cache overhead included, the same yardstick for every proxy above. zoxy's own memory figure elsewhere on this site (the homepage) is a different number by design: the exact total zoxy's own startup banner reports, its fixed pools summed at parse time rather than a sampled peak — the two measure different things, not a change in behavior.

Successful req/s vs offered

open-loop ramp; dashed gray = perfect keep-up

Proxy CPU vs offered

container cores (cAdvisor), mapped onto the offered axis

p99 latency vs offered (while keeping up)

per-window tail (log scale); each line stops where that proxy stops keeping up

Load shed vs offered

offered load the proxy couldn't serve (1 − achieved/offered); under overload most proxies fall behind rather than reject

Latency distribution · HdrHistogram (at 2,000 req/s reference load)

zoxy

latency by percentile — at 2,000 req/s reference load · raw zoxy.hgrm = whole run

haproxy

latency by percentile — at 2,000 req/s reference load · raw haproxy.hgrm = whole run

nginx

latency by percentile — at 2,000 req/s reference load · raw nginx.hgrm = whole run

pingora

latency by percentile — at 2,000 req/s reference load · raw pingora.hgrm = whole run

envoy

latency by percentile — at 2,000 req/s reference load · raw envoy.hgrm = whole run

run 20260813-190516 · generated