DGX Spark 如何避免 OOM 當機

# 參考 https://forums.developer.nvidia.com/t/mitigating-oom-system-freezes-on-uma-based-single-board-computers/362769

# 另外可參考 DGX Spark 之溫度 https://yingrenn.blogspot.com/2026/02/dgx-spark.html

# 安裝輕量級的 Dropbear SSH

$ sudo apt update && sudo apt install dropbear

$ sudo vi /etc/default/dropbear

NO_START=0

DROPBEAR_PORT=2222

$ sudo systemctl enable dropbear

$ sudo systemctl start dropbear

# Standard connection (OpenSSH)

$ ssh spark@<your-ip>

# Emergency connection (Dropbear)

$ ssh spark@<your-ip> -p 2222

# 安裝 earlyoom

$ sudo apt update

$ sudo apt install earlyoom

$ sudo vi /etc/default/earlyoom

EARLYOOM_ARGS="-m 5 -s 10 --avoid 'pipewire|wireplumber|systemd|ssh|journald' --prefer 'vllm|python|triton'"

# This tells earlyoom to intervene when RAM is under 5% AND Swap is under 10%. 

# It will aggressively target vllm or Python scripts over other processes

$ sudo EDITOR=vi systemctl edit earlyoom

### Editing /etc/systemd/system/earlyoom.service.d/override.conf

### Anything between here and the comment below will become the contents of the drop-in file

[Service]

LimitMEMLOCK=infinity

CapabilityBoundingSet=CAP_IPC_LOCK CAP_SYS_NICE CAP_KILL

AmbientCapabilities=CAP_IPC_LOCK CAP_SYS_NICE CAP_KILL

MemoryLock=infinity

OOMScoreAdjust=-1000

### Edits below this comment will be discarded

### /usr/lib/systemd/system/earlyoom.service

# [Unit]

# Description=Early OOM Daemon

# Documentation=man:earlyoom(1) https://github.com/rfjakob/earlyoom

# 

# [Service]

# EnvironmentFile=-/etc/default/earlyoom

# ExecStart=/usr/bin/earlyoom $EARLYOOM_ARGS

# # Run as an unprivileged user with random user id

# DynamicUser=true

# # Allow killing processes and calling mlockall()

# AmbientCapabilities=CAP_KILL CAP_IPC_LOCK

# # We don't need write access anywhere

# ProtectSystem=strict

# # We don't need /home at all, make it inaccessible

# ProtectHome=true

# # earlyoom never exits on it's own, so have systemd

# # restart it should it get killed for some reason.

# Restart=always

# # set memory limits and max tasks number

# TasksMax=10

# MemoryMax=50M

# 

# [Install]

# WantedBy=multi-user.target

$ sudo systemctl daemon-reload

$ sudo systemctl restart earlyoom

$ sudo systemctl status earlyoom

# 查詢 log

$ journalctl -u earlyoom -f

$ cat /etc/systemd/system/earlyoom.service.d/override.conf 

[Service]

LimitMEMLOCK=infinity

CapabilityBoundingSet=CAP_IPC_LOCK CAP_SYS_NICE CAP_KILL

AmbientCapabilities=CAP_IPC_LOCK CAP_SYS_NICE CAP_KILL

MemoryLock=infinity

OOMScoreAdjust=-1000

$ ps aux |grep earlyoom

earlyoom   80791  0.0  0.0   2288  1688 ?        SLs  11:55   0:00 /usr/bin/earlyoom -m 5 -s 10 --avoid pipewire|wireplumber|systemd|ssh|journald --prefer vllm|python|triton

$ cat /proc/$(pgrep earlyoom)/oom_score_adj

-1000