mindstudio_monitor Interfaces

mindstudio_monitor Module

Provides inter-process communication (IPC) interfaces and the capability of independently controlling MSPTI Monitor to collect and obtain profile data.

1. IPC control channel: The profiler backend obtains the profiler configuration from the dynolog daemon.
2. IPC data channel: MSPTI Monitor sends profile data to the dynolog daemon.
3. Lightweight profile data collection

• Starts or stops MSPTI Monitor to collect data.
• Obtains the profile data collected by MSPTI Monitor online.
• Exports the profile data collected by MSPTI Monitor to the local PC in Excel format.

PyDynamicMonitorProxy API

Communicates with the dynolog daemon through IPC, and sends registration requests and profiler configuration parameters to the dynolog daemon. You do not need to directly call this API.

• init_dyno sends registration requests the dynolog daemon.
  • Input: npu_id(int)
  • Return: None
• poll_dyno obtains profiler control parameters from the dynolog daemon.
  • Input: None
  • Return: str, control parameter.
• enable_dyno_npu_monitor enables MSPTI monitoring.
  • Input: cfg_map(Dict[str,str]) parameter configuration
  • Return: None
• finalize_dyno releases resources and threads in mindstudio_monitor.
  • Input: None
  • Return: None
• update_profiler_status reports the profiler status.
  • Input: status(Dict[str,str])
  • Return: None

Monitor Feature APIs

ActivityKind Enumeration Class

This enumeration class defines the types of the data can be collected by MSPTI Monitor and is used to configure the monitor module. Each enumerated value corresponds to a data type.

• ActivityKind.Marker: collects mstx dotting data and returns the marker data structure.
• ActivityKind.Kernel: collects the time consumption data of compute operators and returns the kernel data structure.
• ActivityKind.Communication: collects the time consumption data of communication operators and returns the communication data structure.
• ActivityKind.API: collects the time consumption data of operator API calls and returns the API data structure.
• ActivityKind.AclAPI: collects the time consumption data of ACL API calls and returns the API data structure.
• ActivityKind.NodeAPI: collects the time consumption data of Node API calls and returns the API data structure.
• ActivityKind.RuntimeAPI: collects the time consumption data of Runtime API calls and returns the API data structure.

Monitor API

• start starts monitor to collect data.
  • Input: kinds(List[ActivityKind]) data type list
  • Return: None
• stop stops monitor to collect data.
  • Input: None
  • Return: None
• get_result obtains the profile data collected by monitor.
  • Input: None
  • Return: Dict[ActivityKind, List[ActivityData]], profile data
• save saves the profile data collected by monitor.
  • Input: file_path(str) file path
  • Return: None

ActivityData Data Structure

Defines the profile data structure collected by monitor.

Marker Structure Fields

• name (str): mstx dotting message content
• sourceKind (str): message source type, which can be Host or Device
• domain (str): name of the domain to which the message belongs
• id (int): message ID
• startNs (int): mstx dotting start time, in nanoseconds
• endNs (int): mstx dotting end time, in nanoseconds
• pid (int): process ID when sourceKind is Host, or 0 when sourceKind is Device
• tid (int): thread ID when sourceKind is Host, or 0 when sourceKind is Device
• deviceId (int): ID of the device to which the marker belongs when sourceKind is Device, or 0 when sourceKind is Host
• streamId (int): ID of the stream to which the marker belongs when sourceKind is Device, or 0 when sourceKind is Host

Kernel Structure Fields

• name (str): name of a compute operator
• startNs (int): operator execution start time, in nanoseconds
• endNs (int): operator execution end time, in nanoseconds
• deviceId (int): ID of the device where the operator is executed
• streamId (int): ID of the stream where the operator is executed
• correlationId (int): operator execution correlation ID, which is used to associate with API data
• type (str): operator type, for example, KERNEL_AICORE, KERNEL_AIVEC, or KERNEL_AICPU

Communication Structure Fields

• name (str): name of a communication operator
• startNs (int): operator execution start time, in nanoseconds
• endNs (int): operator execution end time, in nanoseconds
• deviceId (int): ID of the device where the operator is executed
• streamId (int): ID of the stream where the operator is executed
• count (int): data volume transmitted by the operator
• dataType (str): data type transmitted by the operator, for example, FP32 or INT8
• commName (str): name of the communicator to which the operator belongs
• algType (str): communication algorithm type of the operator, for example, RING or MESH
• correlationId (int): operator execution correlation ID, which is used to associate with API data

API Structure Fields

• name (str): API name
• startNs (int): API call start time, in nanoseconds
• endNs (int): API call end time, in nanoseconds
• pid (int): ID of the process that calls the API
• tid (int): ID of the thread that calls the API
• correlationId (int): API call correlation ID, which is used to associate with kernel/communication data

Installation

For details about how to install the mindstudio_monitor module, see msMonitor Installation Guide.