CPU / memory / disk / GPU stress tests with a small web UI.
| Test | Tool | Notes |
|---|---|---|
| CPU | sysbench cpu |
Pick threads + duration. |
| Memory | sysbench memory |
1M block writes; default size auto-tuned to 2× RAM. |
| Disk | fio |
Random 4K IOPS profile + sequential 1M throughput. |
| GPU | gpu-fryer | fp16 matmul on tensor cores; NVIDIA only. |
Each card has its own Run button; tests run concurrently if you fire several at once. Live output streams into each card's panel. State is in-memory only - restarting the container clears job history.
1. Inside the Reefy app catalog (zero setup). Boot a machine into Reefy OS, adopt it on reefy.ai, open the device's Install app menu, pick Bench. A few clicks and you're in. Reefy provisions the container, attaches the GPU via CDI on hosts that have one, sets up the tunnel + access link, and hands you back a URL.
2. Standalone on any Docker host. The image is published publicly
to ghcr.io/reefyai/reefy-bench:latest - run it anywhere. GPU access
goes through CDI (the same mechanism Reefy itself uses), so you need
the NVIDIA Container Toolkit configured to generate a CDI spec at
/etc/cdi/nvidia.yaml (see NVIDIA's CDI guide).
Once that's in place:
# GPU machine (NVIDIA CDI configured)
docker run --rm -p 8500:8500 \
--device nvidia.com/gpu=all \
ghcr.io/reefyai/reefy-bench:latest
# CPU / disk / mem only - no GPU flag needed
docker run --rm -p 8500:8500 ghcr.io/reefyai/reefy-bench:latest
open http://localhost:8500Without --device nvidia.com/gpu=all the GPU card auto-hides; the
other three tests keep working.
No auth: inside the Reefy app catalog the per-device tunnel + service token handle access control; standalone you're responsible for not exposing port 8500 to the open internet.