Ubuntu 26.04 + Angie beats CentMinMod on the same EPYC host — XenForo 2.3.13, PHP 8.2–8.5
Same AMD EPYC 4245P (6C/12T, 30 GiB, 2×894 GB NVMe), same XenForo dataset, same wrk 4.1.0 method as the AlmaLinux 10 + CentMinMod run.
TL;DR
| Stack (PHP 8.2–8.5 avg) | Metric | CentMinMod | Ubuntu + Angie | Delta |
|---|
| 3 — Redis + page cache | Guest cached | 157,957 | 390,232 | +147.1 % |
| 4 — + Anubis | Guest cached | 59,958 | 68,725 | +14.6 % |
| 1 — Plain | Guest bypass | 596 | 775 | +30.1 % |
| 3 — Redis + page cache | Member home | 542 | 747 | +37.9 % |
| 3 — Redis + page cache | Member thread | 694 | 838 | +20.7 % |
Every comparable cell is above CentMinMod. The largest single difference is a TLS setting:
ssl_buffer_size 1400 split each cached page into ~11 TLS records,
16k sends it as one. On the same page and TLS stack that alone moved the cached guest tier from
151,438 to 442,765 req/s.
Environment
| Spec | AlmaLinux 10 + CentMinMod | Ubuntu 26.04 + Angie |
|---|
| CPU / RAM | AMD EPYC 4245P, 6C/12T, 30 GiB | Same host (30,663 MiB) |
| Disk / OS | 2× 894 GB Samsung NVMe, AlmaLinux 10.2, kernel 6.12 | 2× 894 GB Samsung NVMe, Ubuntu 26.04.1, kernel 7.0.0-43 |
| Web server | Nginx 1.31.8, 6 workers, 300k connections | Angie 1.12.4, 8 workers, 32k connections |
| PHP | 8.2.35, 8.3.36, 8.4.27, 8.5.12 (compiled, -O3 -march=native) | 8.2.35, 8.3.36, 8.4.27, 8.5.12 (Sury) |
| PHP-FPM | ondemand, 30 children, TCP 127.0.0.1:9000 | static, 12 children, Unix socket |
| OPcache | 384 MB, JIT disabled | 512 MB, JIT tracing 64 MB, huge pages |
| Database | MariaDB 10.11.19, 8 GB pool, 600 conns | MariaDB 11.8.6, 1.9 GB pool, 200 conns |
| Cache | Redis 8.6.8 + phpredis | Valkey 9.0.4 + phpredis |
| Anubis | 1.28.2 (minimal Nginx terminator) | 1.28.2 (Angie terminator, Unix sockets, static shadow) |
| XenForo | 2.3.13 + 5 add-ons | 2.3.13 + 5 add-ons |
Dataset and method
- 20,001 users, 10,000 threads, 100,400 posts, 47 nodes (CentMinMod run: 101,134 posts, 51 nodes) — Hampel Seeder 1.0.0b11, 2-year span.
wrk -t4 -c100 -d10s, 2 s warm-up, loopback HTTPS against xf.test.local (self-signed).
- Guest cached:
GET / no cookies, Accept-Encoding: gzip (the gzip cache entry; the identity page is ~106 KB and is bandwidth-bound on both platforms).
- Guest bypass:
Cookie: xf_session=bench-bypass. Member: fresh XenForo session cookie, GET /.
- Member thread: the median 9-post thread (129 KB); the first 18-post thread (191 KB) is listed separately in the caveats.
- Page cache cleared per stack; Valkey DB 1 flushed when the Redis cache was toggled; rate limits disabled for the run.
- Stack 4 cached is measured with
ab -k -c100 because wrk 4.1.0 under-reads body bytes from the Go proxy on this host (it reported 4 KB per 15.6 KB response); ab and oha agree with curl. Stack 4 dynamic tiers use oha for the same reason.
Results — averages across PHP 8.2–8.5
| Stack | Platform | Guest cached | Guest bypass | Member home | Member thread |
|---|
| 1 — Plain | CentMinMod | 599 | 596 | 530 | 705 |
| 1 — Plain | Ubuntu | 688 (+14.7 %) | 775 (+30.1 %) | 663 (+25.1 %) | 767 (+8.7 %) |
| 2 — Redis | CentMinMod | 600 | 595 | 542 | 697 |
| 2 — Redis | Ubuntu | 705 (+17.5 %) | 813 (+36.6 %) | 745 (+37.6 %) | 837 (+20.0 %) |
| 3 — Redis + page cache | CentMinMod | 157,957 | 595 | 542 | 694 |
| 3 — Redis + page cache | Ubuntu | 390,232 (+147.1 %) | 819 (+37.8 %) | 747 (+37.9 %) | 838 (+20.7 %) |
| 4 — + Anubis | CentMinMod | 59,958 | 577 | 525 | 666 |
| 4 — + Anubis | Ubuntu | 68,725 (+14.6 %) | 709 (+22.9 %) | 654 (+24.6 %) | 690 (+3.5 %) |
Cached guest tier per PHP version
| PHP | CentMinMod S3 | Ubuntu S3 | CentMinMod S4 | Ubuntu S4 |
|---|
| 8.2 | 146,058 | 389,514 | 60,223 | 70,946 |
| 8.3 | 170,437 | 392,053 | 60,416 | 69,209 |
| 8.4 | 148,069 | 390,677 | 59,821 | 66,145 |
| 8.5 | 167,264 | 388,684 | 59,371 | 68,600 |
Dynamic tiers at PHP 8.5
| Stack | Bypass | Member home | Member thread (9 posts) | Member thread (18 posts) |
|---|
| CentMinMod 1 | 593 | 526 | 699 | — |
| Ubuntu 1 | 791 | 667 | 761 | 613 |
| CentMinMod 3 | 589 | 539 | 691 | — |
| Ubuntu 3 | 841 | 770 | 863 | 678 |
| CentMinMod 4 | 573 | 521 | 664 | — |
| Ubuntu 4 | 722 | 668 | 699 | 566 |
What made the difference
- 16k TLS records.
ssl_buffer_size 1400 fits one MTU but turns a 15.6 KB cached page into ~11 TLS records (and ~11 write syscalls). 16k sends it as one: 151,438 → 442,765 req/s on the same page and TLS stack. This is the whole Stack 3 gap and more.
- 8 Angie workers on 12 CPUs.
worker_processes auto (12) spent cores on TLS contention: 348–352k cached req/s. Two workers per three CPUs (8) gives 389–390k, with member/thread tiers unchanged (764–767 / 840–843 req/s).
- Anubis with the Go collector out of the way.
GOGC=off with GOMEMLIMIT=2GiB (the limit still bounds the heap; RSS < 2 GiB): 51,949 → 57,082 req/s through the filter. The static-asset shadow keeps CSS/JS/images off the proxy on top of that.
- OPcache huge pages.
vm.nr_hugepages sized from RAM plus opcache.huge_code_pages=1: +3.5 % on member/thread (808 → 835, 645 → 670 req/s).
- The rest of the Ubuntu profile (already present): tracing JIT,
pm=static one worker per CPU, gzip stored in the FastCGI cache, aio off + brotli off on cache hits, Unix sockets end to end, MariaDB pool-of-threads, XF data cache in Valkey.
- Sweeps ruled out the usual suspects: 24/32 PHP workers were slower than 12 (536/534 vs 548 req/s), JIT 128 MB and altered hotness thresholds were noise,
validate_timestamps=0 was noise, Anubis GOMAXPROCS 12/16 was within noise of 8.
Why CentMinMod was competitive
- PHP is compiled with
-O3 -march=native (CentMinMod) vs generic Sury packages: roughly 5–10 % per-request advantage.
- 30
ondemand PHP-FPM children over TCP hide database waits better than one static worker per CPU — that is why the 18-post thread page is still slightly behind, while member-home is 25–38 % ahead on Ubuntu.
- Nginx 1.31.8 with newer OpenSSL and
sendfile_max_chunk 2m, tcp_nodelay, open_file_cache.
- Stock CentMinMod has no FastCGI page cache and no Anubis; those were added for the benchmark. Its 6 Nginx workers and JIT-disabled PHP leave performance unused.
Caveats
- CentMinMod’s page-cache config was a minimal addition, not its stock vhost; Ubuntu was tuned end to end. This measures the platforms, not identical configs.
- Thread-page throughput is page-size sensitive. The CentMinMod dataset averages 10.1 posts/thread, so the 9-post page is the fair comparison; the 18-post page (191 KB) measures 613–678 req/s on Ubuntu Stacks 1–3 and 566 on Stack 4.
- Single 10 s run per cell, except Stack 4 cached (2–4 runs, 66–74k, median reported). Cache-hit variance on this host is ±5 %.
- Stack 4 cached uses
ab and dynamic tiers use oha because wrk 4.1.0 misreads Go-proxy body lengths on this host; CentMinMod numbers are wrk. Cached tiers are server-bound and the tools agree within 1–2 % on Stacks 1–3.
- PHP 8.6 RC5 is not in the matrix (CentMinMod EL10 supports up to 8.5); an earlier audit measured ~+6 % thread throughput on 8.6 RC5, offset by missing redis/imagick packages.
Configuration highlights
Code:
# /tls/ssl_engine.conf (the 3x cached-tier change)
ssl_buffer_size 16k;
# angie.conf (2/3 of CPUs; 8 on a 12-thread host)
worker_processes 8;
worker_cpu_affinity auto;
worker_rlimit_nofile 131072;
worker_connections 32768;
# Anubis unit (>=16 GB hosts; GOMEMLIMIT still bounds the heap)
Environment=GOGC=off
Environment=GOMEMLIMIT=2048MiB
Environment=GOMAXPROCS=8
# kernel + OPcache
vm.nr_hugepages = 479
opcache.huge_code_pages = 1
opcache.jit = tracing
opcache.jit_buffer_size = 64M
# PHP-FPM (12 threads)
pm = static
pm.max_children = 12
listen = /run/php/php-fpm.sock
Conclusion
On the same EPYC 4245P the Ubuntu 26.04 + Angie stack is now ahead on every measured tier:
+147 % cached guest throughput (390,232 vs 157,957 req/s),
+14.6 % with Anubis (68,725 vs 59,958),
+38 % member home (747 vs 542) and
+21 % member thread (838 vs 694). The decisive change was not a new service — it was one TLS directive, one worker-count rule, a Go runtime profile and PHP huge pages, all config-only.
Ubuntu 26.04.1 · Angie 1.12.4 · PHP 8.2–8.5 · MariaDB 11.8.6 · Valkey 9.0.4 · Anubis 1.28.2 · XenForo 2.3.13