diff --git a/40-inventory/hosts.yml b/40-inventory/hosts.yml index 8c1aac9..d2ccf3d 100644 --- a/40-inventory/hosts.yml +++ b/40-inventory/hosts.yml @@ -708,22 +708,25 @@ name: jellyfin notes: 'Home LAN media server (Jellyfin :8096, native systemd service, no docker). 16 GB RAM + 1.4 GB swap, swappiness 60. 20 cores (Xeon). RESOLVED (verified 2026-10-26 - by rei): the daily OOM-kill cycle diagnosed 2026-10-07 is stopped — MemoryHigh=8G - + MemoryMax=10G applied 2026-10-08 21:23 UTC via ''systemctl set-property'' (drop-ins - /etc/systemd/system.control/jellyfin.service.d/50-MemoryHigh.conf, 50-MemoryMax.conf); - kernel OOM in the last 10 days: zero (journalctl -k and dmesg clean, cgroup memory.events - oom_kill=0). Jellyfin now runs near the high watermark (memory.current ~7.8G vs - high 8G, >1M throttling events — normal cgroup behavior). vm.swappiness and HW-transcode - suggestions remain NOT applied. NEW KNOWN BEHAVIOR (2026-10-26): scheduled ''Validating - media library'' runs 2x/day (~11:03 and ~23:04 UTC, drift ~2 min/day, 14 runs - in 7 days) saturates CPU/IO (load1 spike to 13.5 on 20 cores, ffprobe/ffmpeg chapter-image - extraction) and can stall /health beyond the panel''s 30 s probe timeout — first - observed 2026-10-26 23:04:24-23:04:54 UTC (episode 30.6 s, ReadTimeout, self-recovered, - no service restart, NRestarts=0). If it recurs or lasts >5 min: SSH triage (journalctl - -u jellyfin, dmesg), consider moving Jellyfin scheduled tasks to night hours or - raising panel probe timeout. GHard panel server id 39. Credentials: gnexus-creds - secret ''jellyfin (192.168.1.153) — SSH ubuntu'' (id 84a01b86-ec1d-41d1-927b-253f9b4c1770), - SSH=sudo same password.' + by rei): daily OOM-kill cycle stopped — MemoryHigh=8G + MemoryMax=10G via systemctl + set-property (drop-ins /etc/systemd/system.control/jellyfin.service.d/); kernel + OOM: zero. vm.swappiness and HW-transcode suggestions remain NOT applied. Trickplay + task manual-only (trigger removed 08/10.10, re-verified still cleared 2026-11-07); + AudioNormalization manual-only since 2026-10-10. KNOWN CLASS (2 episodes): scheduled + ''Validating media library'' / ''Сканирование медиатеки'' (id 7738148f, drifts + ~2 min/day) → ffprobe/ffmpeg preview storm over NFS; 2026-10-26 23:04 ReadTimeout + 30.6 s; 2026-11-07: reboot ~05:20 UTC (home network outage) then 17:21 UTC scan + — service_down ReadTimeout 30.5 s (17:22:31→17:23:01) + cpu_high 91% / load1 peak + 77 / 8–12 parallel ffmpeg thumbnail extraction over new serials batch on NetTera8 + until ~18:08+, health stayed 200 locally, task 95.8% left to finish, self-recovered, + no intervention. Escalation rule: episode within 24 h of previous, or down >5 + min → notify + extended triage; single ~30 s episode in validation window — log + only. NetTera8 (192.168.1.131) 100% full (df from host: 35 MB free of 7.3 TB; + panel 95% masked by ext4-reserved) — metadata writes fail, cleanup/expand = owner + decision. Scan/validate triggers → night hours = owner decision. GHard panel server + id 39; health checker service id 7. Credentials: gnexus-creds secret ''jellyfin + (192.168.1.153) — SSH ubuntu'' (id 84a01b86-ec1d-41d1-927b-253f9b4c1770), SSH=sudo + same password.' owner: gmikcon purpose: Home media server (Jellyfin) ram_gb: 16 @@ -942,3 +945,46 @@ status: active type: vm user: ubuntu +- id: gmikcon-pc + active_services: + - ollama.service + - sshd.service + addresses: + lan: + - 192.168.1.129 + cpu_model: AMD Ryzen 5 3600 6-Core + docs: ../10-systems/hosts/gmikcon-pc.md + environment: home + gpu: NVIDIA GeForce RTX 3090 24GB (driver 615.78.08) + last_reviewed: '2026-11-08' + name: gmikcon-pc + notes: 'Личный десктоп владельца (Arch Linux, zsh, user gmikcon; sudo = SSH-пароль, + gnexus-creds secret ''gmikcon PC (192.168.1.129) — SSH gmikcon'' id cfc8ee47). + Порты 22 (SSH, есть fail2ban-подобный cooldown после неудачных попыток) и 11434 + (Ollama, открыт в LAN). MAC 74:56:3c:b4:cb:18. RAM 62 GiB. 2026-11-08 оптимизация + Ollama под qwen3.6:35b-128k (архитектура qwen35moe: 256 экспертов/8 активных, + гибридное внимание full_attention_interval=4, native ctx 262144, GQA 2 KV-heads, + capabilities completion+vision+tools+thinking) под 3 параллельных запроса x 128k + ctx + 3 ГБ VRAM-резерв под ОС. Ключевой шаг: обновление Ollama 0.30.11 -> 0.40.15 + (старый планировщик не давал parallel для recurrent/moe GGUF; 0.40.15 снял блок), + num_gpu удалён из params манифеста модели (блокировал fit: ''n_gpu_layers already + set by user'') — создана отдельная модель qwen3.6:35b-128k-fit (оригинал не тронут). + Итоговый systemd override /etc/systemd/system/ollama.service.d/override.conf (бэкап + override.conf.bak-rei-20261108): OLLAMA_HOST=0.0.0.0:11434, FLASH_ATTENTION=1, + KV_CACHE_TYPE=q8_0, NUM_PARALLEL=3, CONTEXT_LENGTH=131072, KEEP_ALIVE=-1, MAX_LOADED_MODELS=1, + OLLAMA_GPU_OVERHEAD=3000000000 (3 GiB резерв), DEBUG убран после отладки, LLAMA_ARG_N_CPU_MOE=19 + (первые 19 из 42 слоёв MoE-эксперты в CPU/RAM, 23 последних на GPU; подбором по + расчёту: free 14572 MiB -> вернуть 10872 -> 23.5 слоя). KV q4 не понадобился: + q8_0 влез (KV-буфер 4080 MiB = 3 слота x ~1360 MiB). Итог: single-stream генерация + 61.3 tok/s (было 38 в режиме все-эксперты-в-RAM), VRAM 19812/24576 MiB (свободно + ~4.7 GiB, резерв 3 GiB соблюдён), RSS llama-server ~10.9 GiB RAM. Приёмочный тест: + 3 одновременных запроса (ru/ru/en), все done=true, 30-38 tok/s на поток, total + wall 6.5 s, VRAM стабильна. Установка 0.40.15: ''Text file busy'' на работающих + CLI-процессах владельца (ollama launch) обойдён через mv-переименование старых + бинарей (rename работает на занятых файлах), старый бинарь сохранён для отката. + Модели на хосте: qwen3.6:35b-128k, qwen3.6:35b (22 GB), gemma4:e2b-it-qat, nomic-embed-text.' + purpose: 'Личный десктоп владельца: локальный LLM-инференс (Ollama, MoE-модели)' + source_of_truth: ssh-host-and-owner-confirmed + status: active + type: workstation + user: gmikcon