Floating point computing performance monitoring apparatus and monitoring method

Abstract

The application discloses a floating point calculation performance monitoring device and a monitoring method thereof, and relates to the technical field of computers, and specifically comprises the following modules: a data acquisition module that receives instructions sent by a client and acquires floating point calculation data signals generated by a CPU in real time; a feature extraction module that comprises an extraction unit and extracts floating point calculation data information features; a data scheduling module that comprises a queue scheduler and a queue manager, wherein the queue scheduler and the queue manager are connected with floating point registers; a data monitoring module that monitors the floating point calculation data information features; and a data recovery module that comprises a first statistical unit for acquiring target floating point calculation data information data sets and a second statistical unit for acquiring called floating point calculation data information data sets. The application monitors multiple data queues through the data monitoring module, improves the monitoring efficiency, and recovers lost floating point calculation data information.

Description

TECHNICAL FIELD

[0001] The present application relates to a floating point calculation performance monitoring device and a monitoring method thereof, and belongs to the technical field of floating point calculation performance monitoring. BACKGROUND

[0002] With the increasing depth of deep learning models, the huge number of parameters makes the size of the model larger and larger, and the amount of calculation also rises. In actual engineering applications, most deep learning models do not need 64-bit floating point precision, even 32-bit. In order to improve the calculation speed and reduce the model occupied space, the BF16 (BFloat16) format floating point number emerges as the times require, and has gradually become a standard of deep learning.

[0003] Due to the limitation of computer storage space and word length, most computers perform scientific calculations (such as numerical nuclear reactor simulation programs) under the Institute of Electrical and Electronics Engineers (IEEE) 754 floating point arithmetic standard. The accumulation effect of rounding errors in floating point calculations may have a serious impact on the calculation results, and even cause disastrous consequences. Therefore, the floating point number performance of the computer needs to be monitored.

[0004] The existing monitoring method for floating point numbers generally includes the following steps: first, real-time acquisition of floating point performance data, second, determination of whether the floating point performance data is abnormal by using a monitoring device, and third, determination of the node position of the abnormal floating point performance data in order to subsequently repair the data. When monitoring the floating point performance data, the usual method is to first set a threshold value, then acquire a floating point performance data to be monitored each time, and compare the size relationship between the floating point performance data value and the set threshold value. For example, it is stipulated that the floating point performance data greater than the threshold value is normal data, otherwise, it is abnormal data. The abnormal data needs to be alarmed. When multiple continuous abnormal data appear, each abnormal data needs to be alarmed. However, the alarm is too frequent and the alarm resources are wasted. Moreover, when collecting the floating point performance data, due to the disconnection of the server and other reasons, the data stored in the server cannot be collected, resulting in the loss of part of the floating point calculation data information, incomplete monitoring data, and low accuracy of the monitoring result. Therefore, it is necessary to provide a floating point calculation performance monitoring device and a monitoring method thereof. SUMMARY

[0005] The technical problem solved by the present application is to provide a floating point calculation performance monitoring device and a monitoring method thereof, which can effectively solve the problems of excessive frequent alarm and excessive waste of alarm resources in the prior art monitoring method for floating point numbers, and the problems of inability to collect the data stored in the server, loss of part of floating point calculation data information, imperfect monitoring data, and low accuracy of monitoring results in the collection of floating point performance data.

[0006] The technical solution adopted by the present application is a floating point calculation performance monitoring device, comprising:

[0007] A data collection module is configured to receive an instruction sent by a client and collect floating point calculation data information generated by a CPU in real time.

[0008] A feature extraction module comprises an extraction unit configured to extract floating point calculation data information features from the floating point calculation data information generated by the data collection module, a preset floating point register, a data queue set in the floating point register, and a plurality of main cells set in the data queue, wherein the floating point calculation data information features are stored in the main cells of the corresponding data queue.

[0009] A data scheduling module comprises a queue scheduler and a queue manager, wherein the queue scheduler and the queue manager are connected with the floating point register, the queue scheduler sends a scheduling request to the queue manager, the queue manager generates scheduling information and transmits the scheduling information to the floating point register, and the corresponding floating point calculation data information features are retrieved.

[0010] A data monitoring module comprises a monitoring unit configured to monitor the floating point calculation data information features, wherein the monitoring unit monitors the floating point calculation data information features, a threshold value is set in the monitoring unit, the floating point calculation data information features are compared with the threshold value to determine whether the floating point calculation data information is abnormal, and the abnormal floating point calculation data information is transmitted to a server database for storage.

[0011] A data recovery module comprises a first statistical unit configured to obtain a target floating point calculation data information dataset and a second statistical unit configured to obtain a retrieved floating point calculation data information dataset, wherein the first statistical unit is connected with the floating point register, the second statistical unit is connected with the queue manager, the target floating point calculation data information dataset in the first statistical unit is compared with the retrieved floating point calculation data information dataset in the second statistical unit to obtain missing information data, and a transmission unit is set to insert the missing information data into the main cell of the queue corresponding thereto.

[0012] Preferably, the floating point calculation performance monitoring device further comprises a data analysis module configured to analyze and process the floating point calculation data information, and transmit the processed data to the feature extraction module through an interface.

[0013] Preferably, the floating-point calculation data information features include a sign bit, an exponent bit, a mantissa bit, and a floating-point calculation duration of the floating-point calculation data information, the data queue has multiple groups, and the multiple groups of data queues store the exponent bits of the floating-point calculation data according to the sign bits, so as to be placed in corresponding main cells, and each group of data queues is provided with two sub-data queues, each of which is provided with multiple sequentially arranged sub-cells, and the mantissa bits of the floating-point calculation data information and the floating-point calculation duration are classified and stored, and each sub-cell is associated with its corresponding main cell.

[0014] Preferably, the monitoring unit is provided with a marking part and a comparison part, the marking part is used for marking each data queue and establishing four sub-index numbers corresponding to the floating-point calculation data information features, and adjacent two of the four sub-index numbers are provided with relevance, and the comparison part is used for judging whether the floating-point calculation duration is greater than a threshold value.

[0015] Preferably, the extraction module establishes a secondary index number corresponding to each of the multiple data queues, the four sub-index numbers of each floating-point calculation data information feature are associated with the secondary index number of the data queue in which the floating-point calculation data information feature is located, the data recovery module is provided with a bitmap, the bitmap includes multiple bitmap units, each bitmap unit has a unique primary index number, each primary index number is associated with a secondary index number, and the missing information data is stored in the corresponding bitmap unit.

[0016] Preferably, the data recovery module further includes a comparison unit and a calculation unit, the comparison unit is used for comparing the target floating-point calculation data information data set in the first statistical unit with the called floating-point calculation data information data set in the second statistical unit, the calculation unit calculates the missing information data according to the comparison information, and calculates the relevance between the missing information data and the bitmap unit in the bitmap.

[0017] Preferably, the floating-point calculation performance monitoring device further includes an information query module, the information query module includes a configuration unit, a range determination unit, a conversion unit, and an export unit, the configuration unit is used for configuring a matching relationship between source data and target query data in a server database and a data verification rule, generating a configuration file, the range determination unit is used for reading the configuration file and corresponding data selection instructions, determining a target data range in the server database, the conversion unit is used for executing a conversion instruction, matching a format of the target data with the target query data, and the export unit is used for executing an export instruction, exporting the target data matched with the target query data.

[0018] A monitoring method of a floating-point calculation performance monitoring device, including the following steps:

[0019] S1, in use, the user sends instructions through the client, the data acquisition module receives the instructions and acquires the floating point calculation data information generated in real time, processes the floating point calculation data information through the data analysis module, and transmits the processed data to the feature extraction module through the interface;

[0020] S2, the extraction unit in the feature extraction module extracts the floating point calculation data information features, and stores the exponent bits, the mantissa bits and the floating point calculation duration of the floating point calculation data information in the corresponding data queue, main cell and sub cell according to the sign bit of the floating point calculation data;

[0021] S3, the queue scheduler in the data scheduling module sends a scheduling request to the queue manager, the queue manager generates scheduling information according to the scheduling request, and transmits the scheduling information to the floating point register, and the data in each data queue in the floating point register is called;

[0022] S4, the monitoring unit in the data monitoring module monitors the floating point calculation data information, marks each data queue through the marking part, and establishes four sub index numbers corresponding to the floating point calculation data information features, judges whether the highest floating point calculation duration in each data queue is greater than the set threshold through the comparison part, if the highest floating point calculation duration is greater than the set threshold, the floating point calculation data information in the data queue is abnormal, the floating point calculation data information corresponding to the highest floating point calculation duration in the data queue is transmitted to the server database and stored as a class of log files, if the highest floating point calculation duration is less than the set threshold, the floating point calculation data information in the data queue is normal, and it is transmitted to the server database and stored as a class of log files;

[0023] S5, when querying the abnormal floating point calculation data information, the configuration unit in the information query module configures the matching relationship between the source data and the target query data in the server database and the data verification rule, generates a configuration file, the range determination unit reads the configuration file, and selects the corresponding data selection instruction, determines the target data range in the server database class file, the conversion unit executes the conversion instruction, matches the format of the target data with the target query data, and the export unit executes the export instruction to export the target data matched with the target query data;

[0024] S6, when the server is disconnected, causing part of the floating point calculation data information to be lost, the target floating point calculation data information data set in the first statistical unit is compared with the called floating point calculation data information data set in the second statistical unit through the comparison unit in the data recovery module, the missing information data is calculated according to the comparison information, the correlation of the missing information data and the bitmap unit in the bitmap is calculated, so that the missing information data is stored in the corresponding bitmap unit, the missing information is transmitted to the secondary index number associated with the bitmap unit and the index number through the transmission unit, and the missing information data is accurately inserted into the main signal element and the sub signal element of the corresponding data queue.

[0025] The beneficial effects of the present application are as follows compared with the prior art:

[0026] 1) The data acquisition module is arranged to receive the client instruction, and the floating point calculation data information generated by the CPU in real time is acquired, the floating point calculation data information is processed through the data analysis module, the features of the floating point calculation data information are extracted through the feature extraction module, the exponent bit, the mantissa bit and the floating point calculation time length feature of the floating point calculation data are classified and stored in the main signal element and the sub signal element in each data queue according to the sign bit, and the floating point calculation time length features of each floating point calculation data information in the sub signal element of each data queue are arranged in order, so that the highest floating point calculation time length is convenient to extract, and subsequent monitoring is facilitated.

[0027] 2) In the present application, the data scheduling module is arranged, the scheduling request is sent to the queue manager through the queue scheduler, the scheduling information is generated by the queue manager and is transmitted to the floating point register, the corresponding floating point calculation data information feature data is called, and the number of scheduling information is uniformly distributed according to the data queue in the process of calling the information data, so that the scheduling load of the queue scheduler is reduced.

[0028] 3) The data monitoring module is set in the application, each data queue is marked by the marking part in the monitoring unit, four sub-index numbers corresponding to the information characteristics of the floating point calculation data are established, whether the longest floating point calculation time in each data queue is greater than the set threshold is judged by the comparison part, if the longest floating point calculation time is greater than the set threshold, the floating point calculation data information corresponding to the longest floating point calculation time in the data queue is abnormal, the floating point calculation data information corresponding to the longest floating point calculation time in the data queue is transmitted to the server database and stored as a kind of log file, the floating point register corresponding data queue deletes the floating point calculation data information corresponding to the longest floating point calculation time, at this time, the longest floating point calculation time in the data queue is compared with the set threshold again, if the longest floating point calculation time is less than the set threshold, the floating point calculation data information in the data queue is normal, and it is transmitted to the server database and stored as a kind of log file, at this time, all the floating point calculation data information in the data queue is transmitted to the server database and stored, and the original floating point calculation data information in the data queue is deleted, by monitoring a plurality of data queues, the monitoring efficiency is improved, it is not necessary to monitor each floating point calculation data information, and resource waste is avoided;

[0029] 4) The data recovery module is set in the application, when the data queue information in the floating point register is retrieved, the target floating point calculation data information dataset is obtained by the first statistical unit, the retrieved floating point calculation data information dataset is obtained by the second statistical unit, and the two are compared by the comparison unit, the missing information data is calculated by the calculation unit according to the comparison information, the correlation between the missing information data and the bitmap unit in the bitmap is calculated, so that the missing information data is stored in the corresponding bitmap unit, the missing information is transmitted to the secondary index number associated with the bitmap unit and matching the index number by the transmission unit, and the missing information data is accurately inserted into the main cell and the sub-cell of the corresponding data queue, when the server is disconnected or other reasons, the lost floating point calculation data information can be recovered, the monitoring data is improved, and the accuracy of the monitoring result is improved. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 It is a system block diagram of the floating point calculation performance monitoring device of the application;

[0031] Figure 2 It is a flow chart of the monitoring method of the floating point calculation performance monitoring device of the application;

[0032] In the figure: 1, data acquisition module; 2, feature extraction module; 3, data scheduling module; 4, data monitoring module; 5, data recovery module; 6, data analysis module; 7, information query module. DETAILED DESCRIPTION

[0033] The application will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0034] Embodiment 1: as shown in a floating point calculation performance monitoring device, comprising: Figures 1-2 Data acquisition module 1: receiving instructions sent by the client, collecting floating point calculation data information generated by the CPU in real time;

[0035] Feature extraction module 2: including an extraction unit, extracting floating point calculation data information features according to floating point calculation data information generated by the data acquisition module 1, pre-setting a floating point register, setting a data queue in the floating point register, and setting a plurality of main cells in the data queue, and storing the floating point calculation data information features in the main cells of the corresponding data queue;

[0036] Among them, a secondary index number corresponding to a plurality of data queues is established, and the four sub-index numbers of each floating point calculation data information feature are associated with the secondary index number of the data queue where it is located (the floating point register sends wireless signals to the data queue, generates a plurality of individual first signaling, and each sub-index number is associated with the secondary index number through the corresponding first signaling);

[0037] Among them, the floating point calculation data information features include the sign bit, the exponent bit, the mantissa bit and the floating point calculation time length of the floating point calculation data information, the data queue has multiple groups, and the multiple groups of data queues store the exponent bit according to the sign bit of the floating point calculation data, so that it is put into the corresponding main cell, and two sub-data queues are set in each group of data queues, and a plurality of sequentially arranged sub-cells are set in each sub-data queue, and the mantissa bit and the floating point calculation time of the floating point calculation data information are classified and stored, and each sub-cell is associated with the corresponding main cell;

[0038] Data scheduling module 3: including a queue scheduler and a queue manager, the queue scheduler and the queue manager and the floating point register are connected with each other, the queue scheduler sends a scheduling request to the queue manager, the queue manager generates scheduling information and transmits it to the floating point register, and the corresponding floating point calculation data information feature data is retrieved;

[0039] Data monitoring module 4: including a monitoring unit for monitoring floating point calculation data information features, the monitoring unit monitors floating point calculation data information features, the monitoring unit is provided with a threshold value, by comparing the floating point calculation data information features with the threshold value, it is judged whether the floating point calculation data information is abnormal, and the abnormal floating point calculation data information is transmitted to the server database for storage;

[0040]

[0041] ​The monitoring unit is provided with a marking part and a comparison part, the marking part is used for marking each data queue and establishing four sub-index numbers corresponding to the information features of the floating-point calculation data, and the correlation between any two adjacent sub-index numbers is established, and the comparison part is used for judging whether the floating-point calculation duration is greater than a threshold value;

[0042] The data recovery module 5 comprises a first statistical unit for obtaining a target floating-point calculation data information data set and a second statistical unit for obtaining a called floating-point calculation data information data set, the first statistical unit is connected with the floating-point register, the second statistical unit is connected with the queue manager, the missing information data is obtained by comparing the target floating-point calculation data information data set in the first statistical unit with the called floating-point calculation data information data set in the second statistical unit, and the transmission unit is set to insert the missing information data into the main cell of the queue where the missing information data corresponds to.

[0043] The data recovery module 5 is provided with a bitmap, the bitmap comprises a plurality of bitmap units, each bitmap unit has a unique first index number, each first index number is associated with a second index number (the bitmap sends a signal instruction to each bitmap unit to establish a plurality of separate second signals, each first index number is associated with a second index number through the second signal), and the missing information data is stored in the corresponding bitmap unit;

[0044] The data recovery module 5 further comprises a comparison unit and a calculation unit, the comparison unit is used for comparing the target floating-point calculation data information data set in the first statistical unit with the called floating-point calculation data information data set in the second statistical unit (the comparison unit extracts the data in the target floating-point calculation data information data set and the called floating-point calculation data information data set, generates a first number pair and a second number pair, the data in the first number pair and the second number pair are arranged in order, the comparison between the first number pair and the second number pair generates different data, which is the missing information data), the calculation unit calculates the missing information data according to the comparison information, and calculates the correlation between the missing information data and the bitmap unit in the bitmap (the calculation unit in the data recovery module receives the signal sent by the computer terminal and generates an instruction, calls the missing information data, generates a corresponding first string, samples and calls the bitmap unit from the bitmap unit, generates a corresponding second string, analyzes the first string and the second string through the training model preset in the calculation unit, and judges the correlation features of the two);

[0045] The data analysis module 6 is further included, and is used for analyzing and processing the floating point calculation data information (the data is parsed by a parsing unit in the data analysis module to generate a data table matched with the characteristics of the floating point calculation data information, and the extraction unit in the feature extraction module extracts the data), and the processed data is transmitted to the feature extraction module 2 through an interface.

[0046] The information query module 7 is further included, and includes a configuration unit, a range determination unit, a conversion unit and an export unit. The configuration unit is used for configuring the matching relationship between the source data and the target query data in the server database and the data verification rule, generating a configuration file, the range determination unit is used for reading the configuration file and corresponding data selection instructions, determining the target data range in the server database, the conversion unit is used for executing conversion instructions to match the format of the target data with the target query data, and the export unit is used for executing export instructions to export the target data matched with the target query data.

[0047] Embodiment 2: as shown in the figure, a monitoring method of a floating point calculation performance monitoring device, comprising the following steps: Figure 2

[0048] S1, in use, the user sends instructions through the client, the data acquisition module 1 accepts the instructions and acquires the floating point calculation data information generated by the CPU in real time, processes the floating point calculation data information through the data analysis module 6, and transmits the processed data to the feature extraction module 2 through an interface;

[0049] S2, the extraction unit in the feature extraction module 2 extracts the characteristics of the floating point calculation data information, and stores the exponent bit, the mantissa bit and the floating point calculation time of the floating point calculation data information in the corresponding data queue, main cell and sub cell according to the sign bit of the floating point calculation data;

[0050] S3, the queue scheduler in the data scheduling module 3 sends a scheduling request to the queue manager, the queue manager generates scheduling information according to the scheduling request, and transmits the scheduling information to the floating point register, and the data in each data queue in the floating point register is called;

[0051] ​S4. The monitoring unit in the data monitoring module 4 monitors the floating-point calculation data. Each data queue is marked by the marking part, and four sub-index numbers corresponding to the characteristics of the floating-point calculation data are established. The comparison part determines whether the highest floating-point calculation time in each data queue is greater than a set threshold (the comparison unit extracts data from the target floating-point calculation data dataset and the retrieved floating-point calculation data dataset, generating a first pair and a second pair, with the data in the first and second pairs arranged in order. The first and second pairs are compared, and the different data generated during the comparison is the missing information data). If the highest floating-point calculation time is greater than the set threshold, the floating-point calculation data in that data queue is abnormal, and the floating-point calculation data corresponding to the highest floating-point calculation time in that data queue is transmitted to the server database and stored as a type I log file. If the highest floating-point calculation time is less than the set threshold, the floating-point calculation data in that data queue is normal, and it is transmitted to the server database and stored as a type II log file. The type II log file stores historical data for easy retrieval later.

[0052] a. When the highest floating-point calculation time exceeds the set threshold, the floating-point calculation data information corresponding to the highest floating-point calculation time in the data queue is transmitted to the server database and stored as a type of log file. The floating-point calculation data information corresponding to the highest floating-point calculation time is deleted from the data queue corresponding to the floating-point register. At this time, the highest floating-point calculation time in the data queue is compared with the set threshold again. The floating-point calculation data information in the data queue is compared according to the above steps.

[0053] b. When the highest floating-point calculation time is less than the set threshold, the floating-point calculation data information in the data queue is normal and is transmitted to the server database to be stored as a type II log file. At this time, all floating-point calculation data information in the data queue is transmitted to the server database for storage, and the original floating-point calculation data information is deleted from the data queue.

[0054] Use steps a and b to determine if there are any abnormalities in the floating-point calculation data.

[0055] S5. When querying abnormal floating-point calculation data, the configuration unit in the information query module 7 configures the matching relationship between the source data and the target query data in the server database and the data verification rules, generates a configuration file, the range determination unit reads the configuration file and, in accordance with the corresponding data selection instructions, determines the target data range in a type of file in the server database, the conversion unit executes the conversion instructions to match the format of the target data with the target query data, and the export unit executes the export instructions to export the target data that matches the target query data.

[0056] S6、When the server is disconnected, causing part of the floating point calculation data information to be lost, the target floating point calculation data information data set in the first statistical unit is compared with the called floating point calculation data information data set in the second statistical unit through a comparison unit in the data recovery module, the missing information data is calculated according to the comparison information, the correlation between the missing information data and the bitmap unit in the bitmap is calculated, the missing information data is stored in the corresponding bitmap unit, the missing information is transmitted to the secondary index number associated with the bitmap unit and matching the index number through a transmission unit, and the missing information data is accurately inserted into the main signal cell and the sub-signal cell of the corresponding data queue.

[0057] The above merely illustrates the present application, but the protection scope of the present application is not limited to this, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application, therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A floating-point computing performance monitoring device, characterized in that: include: Data acquisition module (1): Receives instructions sent by the client and collects floating-point calculation data information generated by the CPU in real time; Feature extraction module (2): includes an extraction unit, which extracts floating-point calculation data information features based on the floating-point calculation data information generated by the data acquisition module (1), presets a floating-point register, sets a data queue in the floating-point register, sets multiple main information elements in the data queue, and stores the floating-point calculation data information features in the main information elements of the corresponding data queue according to their classification. Data scheduling module (3): includes a queue scheduler and a queue manager. The queue scheduler, the queue manager, and the floating-point register are interconnected. The queue scheduler sends a scheduling request to the queue manager. The queue manager generates scheduling information and transmits it to the floating-point register to retrieve the corresponding floating-point calculation data information feature data. Data monitoring module (4): includes a monitoring unit for monitoring the characteristics of floating-point calculation data information. The monitoring unit monitors the characteristics of floating-point calculation data information. The monitoring unit is set with a threshold. By comparing the characteristics of floating-point calculation data information with the threshold, it is determined whether the floating-point calculation data information is abnormal, and the abnormal floating-point calculation data information is transmitted to the server database for storage. Data recovery module (5): includes a first statistical unit for acquiring the target floating-point calculation data information dataset and a second statistical unit for acquiring the retrieved floating-point calculation data information dataset. The first statistical unit is connected to the floating-point register, and the second statistical unit is connected to the queue manager. By comparing the target floating-point calculation data information dataset in the first statistical unit with the retrieved floating-point calculation data information dataset in the second statistical unit, the missing information data is obtained. A transmission unit is set up to insert the missing information data into the main information cell of the queue in which it is located. The monitoring unit is provided with a marking part and a comparison part. The marking part is used to mark each data queue and establish four sub-index numbers corresponding to the floating-point calculation data information features. There is a correlation between any two adjacent sub-index numbers. The comparison part is used to determine whether the floating-point calculation time is greater than a threshold. For multiple data queues, a corresponding secondary index number is established. The four sub-index numbers of each floating-point calculation data information feature are associated with the secondary index number of the data queue in which it is located. The data recovery module (5) is provided with a bitmap. The bitmap includes multiple bitmap units. Each bitmap unit has a unique primary index number. Each primary index number is associated with a secondary index number. The missing information data is stored in the corresponding bitmap unit.

2. The floating-point calculation performance monitoring device according to claim 1, characterized in that: It also includes a data analysis module (6), which is used to analyze and process floating-point calculation data information and transmit the processed data to the feature extraction module (2) through an interface.

3. The floating-point calculation performance monitoring device according to claim 2, characterized in that: The floating-point calculation data information features include the sign bit, exponent bit, mantissa bit, and floating-point calculation time of the floating-point calculation data information. There are multiple groups of data queues. The multiple groups of data queues group and store the exponent bit of the floating-point calculation data according to the sign bit, and put it into the corresponding main information cell. Each group of data queues has two sub-data queues. Each sub-data queue has multiple sequentially arranged sub-information cells. The mantissa bit and floating-point calculation time of the floating-point calculation data information are classified and stored. Each sub-information cell is associated with its corresponding main information cell.

4. The floating-point calculation performance monitoring device according to claim 1, characterized in that: The data recovery module (5) further includes a comparison unit and a calculation unit. The comparison unit is used to compare the target floating-point calculation data information dataset in the first statistical unit with the retrieved floating-point calculation data information dataset in the second statistical unit. The calculation unit calculates the missing information data based on the comparison information and calculates the correlation between the missing information data and the bitmap unit in the bitmap.

5. The floating-point calculation performance monitoring device according to claim 4, characterized in that: It also includes an information query module (7), which includes a configuration unit, a range determination unit, a conversion unit, and an export unit. The configuration unit is used to configure the matching relationship between the source data and the target query data in the server database and the data verification rules, and generate a configuration file. The range determination unit is used to read the configuration file and, in accordance with the corresponding data selection instructions, determine the target data range in the server database. The conversion unit is used to execute conversion instructions to match the format of the target data with the target query data. The export unit is used to execute export instructions to export the target data that matches the target query data.

6. A monitoring method for a floating-point computing performance monitoring device, characterized in that: Includes the following steps: S1. When in use, the user sends an instruction through the client. The data acquisition module (1) receives the instruction and collects the floating-point calculation data information generated by the CPU in real time. The data analysis module (6) processes the floating-point calculation data information and transmits the processed data to the feature extraction module (2) through the interface. S2. The features of floating-point calculation data information are extracted by the extraction unit in the feature extraction module (2), and the exponent, mantissa and floating-point calculation duration of the floating-point calculation data information are classified and stored in the corresponding data queue, main information element and sub-information element according to the sign bit of the floating-point calculation data. S3. Send a scheduling request to the queue manager through the queue scheduler in the data scheduling module (3). The queue manager generates scheduling information according to the scheduling request and transmits the scheduling information to the floating-point register to retrieve the data in each data queue in the floating-point register. S4. The floating-point calculation data information is monitored by the monitoring unit in the data monitoring module (4). Each data queue is marked by the marking part, and four sub-index numbers corresponding to the characteristics of the floating-point calculation data information are established. The comparison part determines whether the highest floating-point calculation time in each data queue is greater than the set threshold. When the highest floating-point calculation time is greater than the set threshold, the floating-point calculation data information in the data queue is abnormal. The floating-point calculation data information corresponding to the highest floating-point calculation time in the data queue is transmitted to the server database and stored as a type of log file. If the highest floating-point calculation time is less than the set threshold, the floating-point calculation data information in the data queue is normal. It is transmitted to the server database and stored as a type of log file. S5. When querying abnormal floating-point calculation data information, the configuration unit in the information query module (7) configures the matching relationship between the source data and the target query data in the server database and the data verification rules, generates a configuration file, the range determination unit reads the configuration file and, in accordance with the corresponding data selection instructions, determines the target data range in a type of file in the server database, the conversion unit executes the conversion instructions to match the format of the target data with the target query data, and the export unit executes the export instructions to export the target data that matches the target query data. S6. When the server disconnects, causing some floating-point calculation data to be lost, the comparison unit in the data recovery module compares the target floating-point calculation data dataset in the first statistical unit with the retrieved floating-point calculation data dataset in the second statistical unit. The calculation unit calculates the missing information data based on the comparison information and calculates the correlation between the missing information data and the bitmap unit in the bitmap, so that the missing information data is stored in the corresponding bitmap unit. Through the transmission unit, the missing information is transmitted to the secondary index number associated with its bitmap unit and matching the index number, and the missing information data is accurately inserted into the main information cell and sub-information cell of the corresponding data queue.

Images