dsh-spark-monitor
English | 简体中文
A DeepSeek Harness plugin that puts a live status monitor for 1–4 DGX Spark nodes at the far right of the composer status bar, and lets you switch their GPU clock tiers from the same button.
Temperature, memory and swap are sampled over SSH (sysfs hwmon and /proc/meminfo); the clock tiers are applied with sudo -n nvidia-smi -lgc on every configured node.
Features
- Aggregate status pills — one pill shows the hottest node, one shows the fullest memory; hovering either lists every node with its own value and sample age.
- Configurable nodes — 1 to 4 nodes, each with its own SSH alias, display label, and per-node temperature / memory switches.
- Configurable clock tiers — any list of MHz values plus
unlimited, applied to all nodes from a single click; the popover highlights the current tier.
- Per-node failure isolation — one unreachable node does not break the others; its pill line shows the error.
- en / zh dictionaries and a native settings form in the Desktop plugin manager.
Requirements
- DeepSeek Harness
>= 0.2.1-rc.2 (Web UI / Desktop profile).
- Node.js
>= 20 (only for local development).
- SSH aliases for every node in
~/.ssh/config, reachable without a password prompt (BatchMode=yes).
- Passwordless sudo on every node for
nvidia-smi (sudo -n nvidia-smi ... must work).
nvidia-smi installed on the nodes (only needed for the clock feature).
Passwordless sudo (clock feature only)
Sampling temperature and memory needs no privileges. Applying a clock tier runs sudo -n nvidia-smi -lgc ..., so the SSH user needs a NOPASSWD rule:
# /etc/sudoers.d/dsh-spark-monitor
your-user ALL=(root) NOPASSWD: /usr/bin/nvidia-smi
Replace your-user with the SSH user. Without the rule the monitoring pills still work; set clock: false in the config to hide the clock button entirely.
Install
This package is distributed from GitHub; it is not published to npm.
Desktop: open the Plugins page, choose Add plugin, and enter github:zer0kara/dsh-spark-monitor (or the absolute path of a local checkout), then reload the page. The Desktop profile is owned by the Electron app, so dsh plugin --profile desktop is not the way to install there.
CLI profiles:
dsh plugin --profile web add github:zer0kara/dsh-spark-monitor
# a local checkout works the same way:
dsh plugin --profile web add file:/absolute/path/to/dsh-spark-monitor
Configuration
The plugin assumes no host names of its own: with no hosts configured it samples nothing and draws nothing in the status bar. Add your nodes in your profile's patch layer (~/.dsh/profiles/<profile>/cordis.patch.yml) or in the Desktop plugin-manager form. Replace node-1 / node-2 below with the SSH aliases from your own ~/.ssh/config:
- id: spark-monitor
name: "@zer0kara/dsh-spark-monitor"
config:
hosts:
- alias: node-1
label: node 1
temperature: true
memory: true
- alias: node-2
label: node 2
temperature: true
memory: true
tiers:
- 1710
- 1950
- 2190
- unlimited
clock: true
intervalSeconds: 300
timeoutSeconds: 20
| Field | Default | Meaning |
|---|
hosts[].alias | — (required) | SSH alias from ~/.ssh/config. Allowed characters: A-Z a-z 0-9 . _ -. With no hosts, nothing is sampled. |
hosts[].label | alias | Display label used in tooltips. |
hosts[].temperature | true | Sample temperatures for this node. |
hosts[].memory | true | Sample memory and swap for this node. |
tiers[] | 1710, 1950, 2190, unlimited | Clock tiers offered by the popover. Numbers are MHz; unlimited runs nvidia-smi -rgc. |
clock | true | Enable the GPU clock tier control. Set false to hide the clock button when nvidia-smi needs a password. |
intervalSeconds | 300 | Host sampling interval (30–86400). |
timeoutSeconds | 20 | Whole SSH command deadline (5–120). |
The Desktop plugin manager renders the same fields as a form.
How it works
- Sampling. The Host half runs one SSH command per node per interval. It reads
hwmon temp*_input files (millidegrees), /proc/meminfo (MemTotal, MemAvailable, SwapTotal, SwapFree), uptime and load. Each node's snapshot is stored independently; ageSeconds is computed by the Host per request.
- Clock tiers. Applying a tier runs
sudo -n sh -c 'nvidia-smi -lgc 0,<MHz> && …' (or -rgc for unlimited) and records the tier in /var/lib/gpu-clock-lock/tier on the node. The record file exists because the driver exposes no query for the locked clock — only set/reset; the optional boot service below writes the same file so the pill can show the boot tier after a reboot.
Optional: persist the clock limit across reboots
nvidia-smi -lgc is a runtime setting, so a reboot clears it. The plugin never installs or edits any service — this is only a companion example for people who want the limit to survive a reboot.
# /etc/systemd/system/gpu-clock-lock.service
[Unit]
Description=Lock GPU clocks for thermal control
[Service]
Type=oneshot
ExecStart=/usr/bin/nvidia-smi -lgc 0,2190
ExecStartPost=/bin/sh -c 'mkdir -p /var/lib/gpu-clock-lock && echo 2190 > /var/lib/gpu-clock-lock/tier && chmod 644 /var/lib/gpu-clock-lock/tier'
RemainAfterExit=yes
[Install]
WantedBy=multi-user.target
Replace both 2190 values with your tier. ExecStartPost only writes the tier record so the status pill can show the boot tier (the driver has no query for the locked clock). Without this service the plugin still works: the tier applies at runtime, and after a reboot the clocks revert and the pill shows — until you pick a tier again.
Limitations
- There is no way to read the actual locked clock back from the driver; the plugin shows the last tier it applied (or the boot service's tier). A manual
nvidia-smi -lgc outside the plugin is invisible.
nvidia-smi -lgc does not survive a reboot; use the optional boot service above to persist it.- The clock feature needs passwordless sudo; without it the apply fails and the error is shown per node.
- SSH aliases must resolve without interactive prompts; use
BatchMode=yes-friendly keys.
Development
pnpm install
node --test
The Host tests use a fake SSH runner (setRemoteRunner) and exercise the registered Fetch routes directly.
License
MIT
