Monitoring method and monitoring device of deep learning processor
Monitoring method and a monitoring device of a deep learning processor, which can effectively monitor running of the deep learning processor, capture an abnormal status in an arbitrary working state, and make the deep learning processor return to a normal state in time is presented. The monitoring method comprises: initializing standard data and storing a standard calculation result corresponding to the standard data, sending the standard data to the deep learning processor, receiving a calculation result returned by the deep learning processor, comparing the received calculation result with the stored standard calculation result, and judging the state of the deep learning processor in accordance with the result of the comparison, judging that the state of the deep learning processor is normal when the compared results are consistent, and judging that the deep learning processor is abnormal when the compared results are inconsistent.
1. A monitoring method of a deep learning processor, comprising:
initializing standard data and storing a standard calculation result corresponding to the standard data;
judging a running state of the deep learning processor;
sending the standard data to the deep learning processor in accordance with the running state of the deep learning processor wherein:
the standard data is periodically sent to the deep learning processor at a time interval when the deep learning processor is in a waiting state, and the standard data is sent to the deep learning processor in intervals of time between calculation tasks when the deep learning processor is in a calculating state;
receiving a calculation result returned by the deep learning processor;
comparing the received calculation result with the stored standard calculation result;
judging that a state of the deep learning processor is normal when the compared results are consistent; and
judging that the state of the deep learning processor is abnormal when the compared results are inconsistent.
2. The method of a deep according to claim 1 , wherein:
the standard data is randomly generated or generated with respect to a specific input and output organization, and the standard calculation result is obtained beforehand by a calculation by a neural network which is the same as a calculation configured on the deep learning processor.
3. A computer system for monitoring a deep learning processor, the computer system comprising memory storing a program, the computer system being configured by execution of the program to:
initialize standard data and storing a standard calculation result corresponding to the standard data;
judge a running state of the deep learning processor;
send the standard data to the deep learning processor in accordance with the running state of the deep learning processor wherein when it is judged that the deep learning processor is in a waiting state, the standard data is periodically sent to the deep learning processor at a time interval, and when it is judged that the deep learning processor is in a calculating state, calculation tasks are separated, and the standard data is sent to the deep learning processor in intervals of the calculation tasks;
receive a calculation result returned by the deep learning processor;
compare the received calculation result with the stored standard calculation result;
judge that a state of the deep learning processor is normal when the compared results are consistent; and
judge that the state of the deep learning processor is abnormal when the compared results are inconsistent.
4. The computer system according to claim 3 , wherein:
the standard data is randomly generated or generated with respect to a specific input and output organization, and the standard calculation result is obtained beforehand by a calculation by a neural network which is the same as a calculation configured on the deep learning processor.