HyperAccel Driver¶
ha.ko is the kernel driver that exposes the HyperAccel LPU to Linux. This documentation is for server operators and administrators who install and operate the driver, and it describes the operational features that the LPU kernel driver provides. The features covered here are telemetry, hang detection, device recovery, firmware update, module parameters, and the driver operation tools; the operational tasks that you often encounter while managing the driver are organized as tutorials with example scripts.
If you are an application developer
How to use the device at the application level (memory allocation/deallocation, kernel execution, use of streams/events) is covered in the Runtime documentation. Applications use the device through the runtime API, and this documentation covers how to operate the driver from a lower-level perspective.
Position in the stack¶
1 2 3 4 5 | |
At a glance¶
| Kernel module | ha.ko |
| Device node | /dev/ha0, /dev/ha1, … one per device |
| Telemetry | Read-only nodes under /sys/class/ha/haN/, and hwmon |
| Module parameters | /sys/module/ha/parameters/ |
| Tools | hyperaccel-smi, ha-hotreset, ha-fw-update, ha-led, ha-wipe |
| Distributed as | deb package (hyperdriver-dkms) |
| Containers | CDI spec (managed by ha-container-toolkit) |
| Version of this documentation | 2.6.0 |
| Status | Beta |
Check the version of the driver installed on the server like this.
1 | |
Features¶
| Feature | What it does | Documents |
|---|---|---|
| Telemetry | Check temperature, power, utilization, memory usage, and the current state of the device | Telemetry, sysfs nodes |
| Hang detection | Response wait ceiling, watchdog, and hang diagnostic records | Hang detection and the watchdog |
| Device recovery | Automatic reset, application reset, and card reset issued from the host server | FATAL and recovery, ha-hotreset |
| Firmware update | Update the boot flash of the card | Firmware update, ha-fw-update |
| Monitoring | Observe the card state as a table or JSON, and connect it to monitoring systems and health checks | hyperaccel-smi, Collecting telemetry with Prometheus, Building a card health check |
| Configuration | Settings such as the watchdog, automatic reset, and telemetry interval | Configuring and using parameters, Parameter reference |
| Multiple devices | Observe the paths between devices | Multiple devices |
| Containers | Node permissions and the CDI spec | Device node permissions and udev rules, Container support |
Trust boundary
Device memory is not isolated between processes that can open the same device node. The default udev rule opens the node to all local users (0666). Separate workloads that must be isolated from each other by device node access permissions. For the paths between multiple devices, see Multiple devices.
Documentation structure¶
| Section | Description |
|---|---|
| Overview and Release notes | Introduction to the driver and changes by version |
| Installation and setup | Supported environments, installation, verifying the installation, upgrading and removal |
| Operations guide | Telemetry, hang detection, recovery, firmware update, multiple devices, driver operation tools, sysfs nodes and kernel log messages, troubleshooting |
| Kernel module parameters | How to change parameters, commonly used settings, definitions of each parameter |
| Examples and tutorials | Monitoring integration, health checks, diagnostic information collection, card recovery, firmware updates for multiple cards |