Performance and sizing
These figures come from a benchmark run on 2026-09-29 against NullWard 0.6.66-beta. Use them to size your deployment. They're relative, per-vCPU figures from a single test machine, not guarantees. See Caveats.
Sizing summary
With every feature turned on (OWASP CRS in block mode, ML, sign-in, rate limiting, IP intelligence, request logging and the other protections), small requests over keep-alive HTTPS:
| 1 vCPU | 2 vCPU | |
|---|---|---|
| Maximum throughput | about 490 req/s | about 1,040 req/s |
| Recommended planning figure (50% of maximum) | about 250 req/s | about 500 req/s |
| Latency at 50% load (p50 / p99) | 3.0 / 11.8 ms | 2.8 / 18.4 ms |
- Plan for about 250 requests per second per vCPU with everything on. At 50% of the maximum, p99 latency was 12–18 ms.
- Request bodies cost much more than small GETs. The OWASP Core Rule Set runs its patterns over every JSON field name and value. A benign 10 KB JSON POST manages about 6 req/s per vCPU with the CRS on. For APIs that take large bodies from trusted clients, use WAF bypass paths or whitelist rules on those paths.
- New TLS connections are expensive. A full TLS handshake per request is the limit for clients that don't reuse connections. Make sure any load balancer in front of NullWard keeps connections alive.
- Throughput roughly doubles from 1 to 2 vCPUs (2.0–2.3x with everything on).
How it was measured
- Machine. AMD Ryzen 9 8945H, Linux under WSL2, Docker. NullWard was pinned to 1 vCPU (one hardware thread, with its SMT sibling idle) or 2 vCPUs (one full physical core, which is how cloud providers count vCPUs). The backend, tunnel connector, SurrealDB, Valkey and load generator each ran on their own cores.
- Validity. NullWard ran at 91–97% of its CPUs at each maximum, while the backend and load generator stayed below about 45%, so the figures measure NullWard. Every response in every run had to be a 2xx from the backend: a single error, 403 or 5xx would have invalidated the run. All 42 runs were valid.
- Paths.
directis client → NullWard → backend.tunnelis client → NullWard → WireGuard tunnel → connector → backend.TLSmeans HTTPS terminated at NullWard (TLS 1.3, ECDSA P-256) with plain HTTP to the backend.plainmeans HTTP throughout. - Requests.
GET /with a response of about 180 bytes, or a benign 10 KB JSONPOST. - Maximum throughput. Closed-loop saturation at 16, 32, 64 and 128 connections, stopping when throughput gained less than 3% or p99 passed 1 s.
- Latency. Open-loop load at 25%, 50% and 75% of that maximum, 10 s each.
Feature levels. Each level adds to the one before. Response caching was off throughout.
| Level | Adds |
|---|---|
| L0 | Plain reverse proxy |
| L1 | OWASP Core Rule Set, block mode |
| L2 | ML WAF, detect mode |
| L3 | Email-code sign-in, with a session cookie checked on every request |
| L4 | Rate limiting (never triggered), IP intelligence, 4xx behaviour scoring |
| L5 | Request logging |
| L6 | HSTS, body size limit, forwarding headers and a cookie rule: everything on |
Results
Throughput over keep-alive connections (requests per second)
| Path | Protocol | Level | 1 vCPU | 2 vCPU |
|---|---|---|---|---|
| direct | TLS | L0 plain proxy | 3,802 | 7,863 |
| direct | TLS | L1 + CRS | 1,428 | 2,176 |
| direct | TLS | L2 + ML | 1,066 | 2,263 |
| direct | TLS | L3 + sign-in | 492 | 1,095 |
| direct | TLS | L4 + protections | 537 | 1,245 |
| direct | TLS | L5 + request log | 506 | 1,085 |
| direct | TLS | L6 everything | 487 | 1,037 |
| direct | plain | L0 | 4,395 | 7,642 |
| direct | plain | L1 | 1,233 | 2,308 |
| direct | plain | L6 | 470 | 1,083 |
| tunnel | TLS | L0 | 2,969 | 4,557\* |
| tunnel | TLS | L1 | 750 | 1,883 |
| tunnel | TLS | L6 | 429 | 948 |
| tunnel | plain | L0 | 3,502 | 4,390\* |
| tunnel | plain | L1 | 1,217 | 2,157 |
| tunnel | plain | L6 | 504 | 1,014 |
| direct | TLS | L0, 10 KB JSON POST | 3,196 | 4,600 |
| direct | TLS | L1, 10 KB JSON POST | 6 | 12 |
| direct | TLS | L6, 10 KB JSON POST | 6 | 14 |
\* The tunnel connector was near its own CPU limit at these points, so they may understate NullWard's capacity.
New TLS connections (one full handshake per request)
| Level | 1 vCPU | 2 vCPU | For comparison: keep-alive at the same level |
|---|---|---|---|
| L0 | 1,374 conn/s | 2,038 conn/s | 3,802 / 7,863 req/s |
| L6 | 293 conn/s | 741 conn/s | 487 / 1,037 req/s |
Latency, direct path with TLS, GET (p50 / p99 in ms)
Measured at 25% · 50% · 75% of each configuration's maximum throughput.
| Level | 1 vCPU | 2 vCPU |
|---|---|---|
| L0 | 0.8 / 5.4 · 1.0 / 180 · 2.6 / 277 | 0.8 / 2.7 · 0.9 / 7.1 · 3.9 / 329 |
| L1 | 1.5 / 4.5 · 12.8 / 1,050 · 2,500 / 3,080 | 2.0 / 190 · 1.9 / 36 · 16.5 / 150 |
| L3 | 3.2 / 6.8 · 3.0 / 13.4 · 3.9 / 113 | 2.9 / 6.8 · 2.6 / 13.3 · 2.7 / 25.9 |
| L6 | 3.2 / 6.7 · 3.0 / 11.8 · 3.5 / 33.3 | 2.9 / 6.5 · 2.8 / 18.4 · 3.3 / 34.7 |
What the numbers show
- The OWASP Core Rule Set is the largest cost. Turning it on removes 60–72% of plain-proxy throughput for small GETs, about 0.45 ms of CPU per request on one vCPU. Bodies cost far more: about 120–160 ms of CPU for a 10 KB JSON body. This cost comes from the rule set's own inspection, not from NullWard's proxy.
- Sign-in is the second-largest cost. Checking a sign-in session on every request roughly halves throughput again, about 1 ms of CPU per request on one vCPU.
- ML, rate limiting, IP intelligence, request logging and the other protections cost little. Each is within measurement noise, except ML at 1 vCPU, which measured up to 25%. Budget 0–25% for the ML WAF.
- TLS costs little on reused connections, a lot on new ones. On keep-alive connections TLS cost 0–15%. A plain proxy handles about 1,374 new TLS connections per second per vCPU, compared with 3,802 requests per second on reused connections.
- Tunnels cost little once features are on. With everything on, routing through a tunnel cost 6–15%, because inspection dominates. For a plain proxy it cost 20–42%. The connector itself used about one CPU core per ~3,000 requests per second for a plain proxy, and about a third of NullWard's CPU with everything on. Size the connector's host accordingly.
Caveats
- Single machine. The benchmark ran on one laptop under WSL2. The same measurement repeated across sessions varied by about ±10–15%, so smaller differences aren't meaningful. That includes the few cells where TLS beat plain HTTP, and L4 measuring above L3.
- Latency at 75% load on one vCPU is unstable. One vCPU has little spare capacity, so short host stalls become long queues: several 1-vCPU cells at 50–75% show p99 of 0.3–3 s. Size from the 25–50% columns and leave headroom.
- The 1-vCPU figures are optimistic. The second hardware thread on the same core was idle, unlike a cloud vCPU with a busy neighbour.
- Not measured: bot detection, malware scanning, the waiting room and maintenance mode, response-cache hits, HTTP/2 and HTTP/3 clients, large responses, many services at once, and clustered nodes.
- Version. These figures are for 0.6.66-beta. Later releases inspect chunked and streamed request bodies too, up to the body inspection limit (default 128 KiB). This makes large uploads cost more than they did in this run, unless their paths are bypassed or excluded.

