JP2022526333A - Data processing methods and equipment, computer devices, recording media, and computer programs - Google Patents

Abstract

The present invention provides a data processing method and apparatus, a computer device, a recording medium, and a computer program used for training a deep learning model, the method performing a first update process on a prefetch quantity of sample data. In response to obtaining the target prefetch quantity and responding that the quantity of sample data currently contained in the prefetch sample data queue has not reached the target prefetch quantity, the new sample data is read and read. Includes storing the prefetch sample data cune with new sample data. [Selection diagram] Fig. 1

Description

<Mutual citation of related applications>
The present invention claims priority for a Chinese patent application with application number 200911403669.4 and filing date December 30, 2019, the entire contents of which are incorporated herein by reference.
The present invention relates to the field of machine learning technology, specifically to data processing methods and devices, computer devices, recording media, and computer programs.

Deep learning models require multiple iterative trainings based on large amounts of sample data. In order to improve the convergence speed in the training process of the machine learning model, it is usually realized by adopting the method of multi-process parallel training. When training is performed on a deep learning model by adopting the method of multi-process parallel training, the calculation task of this training is executed, and at the same time, the training data required for the next training is prepared in advance in each parallel process. It will be read. Each parallel process must perform communication and data synchronization between all processes after performing that round of training, and if one process reads the training data used for the next training too slowly. , The entire training process is delayed, leading to reduced training efficiency.

The embodiments of the present invention provide at least a data processing method and apparatus.

According to a first aspect, embodiments of the invention provide a data processing method that is applied to training a deep learning model that includes one or more processes, the method of which is said one or more. For one target process in the process, the first update process is executed for the prefetch quantity of sample data to obtain the target prefetch quantity, and it is currently included in the prefetch sample data queue corresponding to the target process. In response to the quantity of the sample data being read not reaching the target prefetch quantity, the new sample data is read and the read new sample data is stored in the prefetch sample data queue.

In this way, the main process performs the first update process on the prefetch quantity to obtain the target prefetch quantity, and when the sample data amount currently contained in the data queue has not reached the target prefetch quantity. To read new sample data from the sample data pool, the main process will perform one iterative training and then complete the reading of the sample data required for the next one iterative training. In most cases in practice, the time it takes for the main process to read the data is often shorter than the time it takes to perform one iterative training, so a sufficient amount of sample data is always stored in the data queue. It can be ensured to meet the use of some subsequent iterative training, and if the main process takes too long to read certain sample data, the sample quantity will not be read in time. It was possible to avoid delays in repetitive training and improve training efficiency.

In one possible embodiment, performing the first update process on the prefetch quantity of the sample data to obtain the target prefetch quantity is currently occupied by the prefetch sample data queue corresponding to the one or more processes. It includes executing the first update process for the prefetch quantity of the sample data to obtain the target prefetch quantity based on the total memory space and the memory usage upper limit threshold.

In this way, the prefetch quantity of the sample data can be dynamically updated based on the total memory space currently occupied by the prefetch sample data queue and the upper limit of memory usage, and the amount of prefetch sample data can be flexibly updated. Can be assigned to meet training requirements.

In one possible embodiment, the first update to the prefetch quantity of the sample data is based on the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes and the upper memory usage threshold. Performing processing to obtain the target prefetch quantity is the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes, the memory usage limit, and the deep layer of the target process. It includes performing a first update process on the prefetch quantity of the sample data to obtain the target prefetch quantity based on the data throughput for performing training on the training model.

In this way, the prefetch quantity is dynamically updated based on the total memory space currently occupied by the prefetch sample data queue, the upper memory usage threshold, and the data throughput when training the deep learning model. Allows the amount of data in the prefetch sample data queue to keep up with the consumption of sample data when the data throughput increases, and the amount of memory occupied by the prefetch sample data queue when the data throughput decreases. Can be reduced as much as possible and excess memory can be used for other tasks, making adjustments more flexible.

In one possible embodiment, the first update to the prefetch quantity of the sample data is based on the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes and the upper limit memory usage threshold. Performing processing to obtain the target prefetch quantity is the prefetch if the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes does not reach the memory usage upper limit threshold. The quantity is increased by the first adjustment step size to obtain the target prefetch quantity, and / or the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes is the memory usage. When the upper limit threshold is reached, the prefetch quantity is reduced by the second adjustment step size to obtain the target prefetch quantity.

In this way, if the total memory space currently occupied by the prefetch sample data queue does not reach the memory usage upper limit threshold, the prefetch sample data is increased as much as possible and the prefetch sample data queue is currently occupied. When the total memory space is reached the upper limit of memory usage, the length of the prefetch sample data queue can be flexibly adjusted by reducing the prefetch quantity of the sample data.

In one possible embodiment, if the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes has not reached the memory usage upper bound threshold, the prefetch quantity is adjusted in the first adjustment step. Increasing by size to obtain the target prefetch quantity means that the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes has not reached the memory usage upper limit threshold and When the data throughput for executing training for the deep learning model of the target process satisfies a predetermined data throughput condition, the prefetch quantity may be increased by the first adjustment step size to obtain the target prefetch quantity. include.

In one possible embodiment, the method is such that the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes does not reach the memory usage upper limit and the data throughput. Further includes reducing the prefetch quantity by a third adjustment step size to obtain the target prefetch quantity if does not meet the predetermined data throughput condition.

In one possible embodiment, the predetermined data throughput condition is such that the current value of the data throughput is larger than the historical value and the current value of the data throughput is larger than the data throughput threshold. Including one, where the historical figures are the average value of the data throughput at the time of multiple historical iterative trainings prior to the current iterative training, or the one-time iterative training prior to the current iterative training. It is a numerical value of the said data throughput at the time.

In one possible embodiment, the method further comprises performing a second update process on the prefetch quantity adjustment step size to obtain a target adjustment step size, wherein the target adjustment step size is: It is used for the next one-time update process of the prefetch quantity.

In one possible embodiment, performing the second update process for the adjustment step size of the prefetch quantity to obtain the target adjustment step size is the case where the prefetch quantity is increased during the first update process. Increasing the prefetch quantity adjustment step size and / or decreasing the prefetch quantity adjustment step size during the first update process includes decreasing the prefetch quantity adjustment step size.

In this way, when the prefetch quantity needs to be increased, increasing the prefetch quantity faster ensures that the sample data stored in the prefetch sample data queue reaches the larger quantity faster and subsequent. Decrease the prefetch quantity more slowly when the requirements for using the training iteration cycle can be met and the prefetch quantity is not too small to delay the model training process, while at the same time the prefetch quantity needs to be reduced. Therefore, it is possible to guarantee that the change in the length of the prefetch sample data queue becomes smoother, and it is possible to avoid the shock of the training process due to the sudden decrease in the quantity of the prefetch sample data.

According to a second aspect, embodiments of the invention further provide a data processing apparatus applicable to training a deep learning model, comprising one or more processes, the apparatus being said one or more. For one target process in the process, the first update module for executing the first update process for the prefetch quantity of sample data to obtain the target prefetch quantity, and the prefetch sample corresponding to the target process. In response to the quantity of sample data currently contained in the data queue not reaching the target prefetch quantity, new sample data is read and the read new sample data is stored in the prefetch sample data queue. It is equipped with a read module for.

According to a third aspect, an embodiment of the present invention further provides a computer device, wherein the computer device comprises a processor, a recording medium, and a bus, the recording medium being feasible by the processor. A machine-readable instruction is recorded, and when the computer device is operated, the processor and the recording medium communicate with each other via a bus, and the machine-readable instruction is executed by the processor. The steps of the first aspect or any possible embodiment of the first aspect are carried out.

According to a fourth aspect, an embodiment of the present invention further provides a computer-readable recording medium, in which a computer program is recorded and the computer program is operated by a processor. , The steps of the first aspect described above or any possible embodiment of the first aspect are performed.

Hereinafter, in order to make the above-mentioned object, features, and advantages of the present invention more clear and understandable, preferred embodiments will be given and described in detail with reference to the drawings.

Hereinafter, in order to more clearly explain the technical solution of the embodiment of the present invention, the drawings required for the embodiment will be briefly introduced. The drawings herein are incorporated herein to form a portion of the specification, which drawings show examples in line with the invention and together with the specification describe technical solutions of the invention. .. It should be understood that the drawings below show only some embodiments of the present invention and should not be considered as limiting the scope. One of ordinary skill in the art can obtain other related drawings based on these drawings without any creative work.
It is a flowchart which shows the data processing method provided by the Example of this invention. It is a schematic diagram which shows the data processing apparatus provided by the Example of this invention. It is a schematic diagram which shows the computer device provided by the Example of this invention.

Hereinafter, in order to further clarify the purpose, technical solution, and advantage of the embodiment of the present invention, the technical solution of the embodiment of the present invention is clarified with reference to the drawings of the embodiment of the present invention. And I will explain it completely. Of course, the examples described are not all examples, but only a part of the present invention. Typically, the components of the embodiments of the invention described and displayed in the drawings herein can be arranged and designed in a variety of different configurations. Accordingly, the detailed description of the embodiments of the invention provided in the following drawings is not intended to limit the scope of the claimed disclosure, but merely represents a selected embodiment of the present disclosure. All other embodiments obtained by one of ordinary skill in the art based on the embodiments of the present invention without creative work shall be included in the scope of protection of the present disclosure.

As a result of the research, when training the deep learning model by adopting the method of multi-process parallel training, the calculation of this training is performed, and at the same time, each process has the training data necessary for the next training in advance. It will be read. Each parallel process should perform communication and data synchronization with other processes after performing this training, and start the next training task after all processes have completed communication and data synchronization. There is. If there is a time delay in the training task of any process here, for example, when pre-reading the training data to be used for the next training, if the reading time exceeds the execution time of this training task, all All training tasks in the process will be delayed in time, leading to reduced training efficiency.

In view of the investigations described above, the present invention provides data processing methods and devices applicable to deep learning model training. In the data processing method, the first update process is executed for the prefetch quantity of the sample data to obtain the target prefetch quantity, and the quantity of the sample data currently contained in the prefetch sample data queue reaches the target prefetch quantity. In response to the absence, the new sample data is read and the read new sample data is stored in the prefetch sample data queue so that the main process performs one iterative training and then the next one. It is possible to ensure that the reading of the sample data required for the iterative training of is already completed. The main process dynamically updates the prefetch quantity to get the target prefetch quantity, and when the amount of sample data currently contained in the data queue does not reach the target prefetch quantity, new sample data from the sample data pool. To read. In most cases, the data queue always stores a sufficient amount of sample data because the main process only takes less time to read new sample data than it takes to perform one iterative training. Can be ensured to meet the use of several subsequent iterative trainings, and if the main process takes too long to read certain sample data, the sample quantity is insufficient and the iterative training is delayed. The training efficiency was improved without the occurrence of.

The deficiencies present in conventional solutions are the result of the inventor's practice and careful study, and therefore the process of finding the above problems and the solutions proposed in the present invention are all books in the process of this disclosure. It should be the inventor's contribution to the disclosure.

Hereinafter, the technical solutions in the present invention will be described clearly and completely with reference to the drawings of the present invention. Of course, the examples described are not all examples, but only a part of the present invention. Typically, the components of the embodiments of the invention described and displayed in the drawings herein can be arranged and designed in a variety of different configurations. Accordingly, the detailed description of the embodiments of the invention provided in the following drawings is not intended to limit the scope of the claimed disclosure, but merely represents a selected embodiment of the present disclosure. All other embodiments obtained by one of ordinary skill in the art based on the embodiments of the present invention without creative work shall be included in the scope of protection of the present disclosure.

It should be noted that in the drawings below, similar signs and letters indicate similar items. Therefore, defining an item in one drawing does not require further definition and analysis in subsequent drawings.

In order to facilitate understanding of the present embodiment, first, the data processing method disclosed by the embodiment of the present invention will be described in detail. The data processing methods provided by the embodiments of the present invention apply to training deep learning models. Its execution subject is generally the main process or subprocess that trains the deep learning model. In some possible implementations, the data processing method may be implemented by a method of calling a computer-readable instruction stored in memory by a processor.

Hereinafter, the data processing method provided by the embodiment of the present invention will be described by taking an example in which the execution subject is an arbitrary main process in at least one main process.

FIG. 1 is a flowchart of a data processing method provided by an embodiment of the present invention, which method comprises steps S101 to S102, wherein the method comprises steps S101 to S102.
In S101, the first update process is executed for the prefetch quantity of the sample data to obtain the target prefetch quantity.
In S102, in response to the fact that the quantity of sample data currently contained in the prefetch sample data queue has not reached the target prefetch quantity, new sample data is read and the read new sample data is used in the prefetch sample data queue. Store inside.

Hereinafter, the above-mentioned S101 to S102 will be described in detail.

I: In S101 described above, when there is one main process, one main process trains the deep learning model, and the main process executes the first update process for the prefetch quantity of the sample data. The target prefetch quantity can be obtained.

If there are multiple main processes, the multiple main processes perform parallel training on the deep learning model, and each main process performs the first update process on the prefetch quantity of the sample data. You can get the prefetch quantity. Here, the prefetch quantities corresponding to different main processes can be different, and the target prefetch quantities corresponding to different main processes can also be different.

Each main process corresponds to one prefetch sample data queue, and the prefetch sample data queue corresponding to any main process contains a plurality of sample data, and each main process has a corresponding prefetch sample. Train a deep learning model based on the sample data stored in the data queue.

The prefetch sample data queue is, for example, a first-in first-out queue. When starting a new iterative training, the main process first reads a set of sample data from the prefetch sample data queue corresponding to the main process. After being read, the set of sample data is deleted from the prefetch sample data queue to reserve a storage position for new sample data.

Note that in one iterative training, the main process will train the deep learning model based on a set of sample data, and one set of sample data will contain at least one sample data. Is. In the embodiment of the present invention, the prefetch quantity can be the quantity of the sample data set.

If the predetermined prefetch quantity update condition is not reached, the main process responds that the quantity of sample data currently contained in the prefetch sample data queue has not reached the current prefetch quantity. Read the sample data.

When the predetermined prefetch quantity update condition is reached, the main process performs the first update process on the current prefetch quantity of the sample data to obtain the target prefetch quantity, and the main process receives the prefetch sample data queue. Read new sample data in response to the fact that the quantity of sample data currently contained in is not reaching the target prefetch quantity. Here, the prefetch quantity and the target prefetch quantity of the sample data may be the same or different.

Specifically, the prefetch quantity update condition includes, for example, one or more of the following a1 to a3.

a1 means that a predetermined update cycle has been reached.

Here, the update cycle is a cycle for updating the prefetch quantity.

The update cycle is, for example, a time cycle. For example, if the update cycle is one hour, the first update process for one prefetch quantity can be triggered every hour.

The update cycle is, for example, a predetermined quantity of iterative training, for example, the main process triggers the first update process for one prefetch quantity for every five iterative trainings performed on the deep learning model. can do. Here, the duration of different update cycles will also be different, as the time required for different repetitive trainings may differ from each other.

A2 is that the data throughput when training the deep learning model based on the sample data is larger than the first threshold value.

Here, the data throughput is used to indicate the speed of sample data processing as the main process trains on the deep learning model. If the data throughput when the main process trains on the deep learning model is greater than the first threshold, it is considered to consume faster for the sample data stored in the prefetch sample data queue. Here, if a smaller prefetch quantity is maintained, the quantity of sample data stored in the prefetch sample data queue may not keep up with the consumption of training in time. Therefore, it is conceivable to increase the quantity of sample data prefetched in the prefetch sample data queue and trigger the first update process for the prefetch quantity of the sample data.

a2.1: Here, for example, the following method is adopted to obtain the data throughput.
In response to reaching the prefetch quantity update condition, at least one target iterative training is determined from multiple historical iterative trainings based on the training progress for the deep training model when the prefetch quantity update condition is reached. Also, based on the quantity of sample data contained in the sample data set used for each target iterative training and the time taken for each target iterative training, the deep learning model is trained based on the sample data. Determine the data throughput when running.

Here, the target iterative training is, for example, at least one iterative training closest to the timing when the prefetch quantity update condition is reached.

For example, one main process has already performed 5 iterative trainings on a deep learning model, and when the prefetch quantity update condition is reached, it has performed a 6th iterative training on the deep learning model. There is. Here, if there is one target repetitive training, the fifth repetitive training can be determined as the target repetitive training. If the time required for the fifth iterative training is 15 minutes, the quantity of sample data used is 64, and the data throughput is, for example, 64/15.

If there are three target repetitive trainings, the third, fourth, and fifth repetitive trainings can be defined as target repetitive trainings. Data throughput when the third, fourth, and fifth iterative trainings take 12 minutes, 14 minutes, and 15 minutes, respectively, and the quantity of sample data used for each iterative training is 64, respectively. Is, for example, 64 × 3 ÷ (12 + 14 + 15), and the unit is pieces / minute.

a2.2: In another embodiment, further, the repetitive training currently being performed is determined as the target repetitive training, and the quantity and duration of the samples already completed during the repetitive training currently being performed are used. Based on this, the data throughput can be determined.

For example, a main process has already performed 5 iterative trainings on a deep learning model, and when the prefetch quantity update condition is reached, it is performing a 6th iterative training on the deep learning model. .. The sixth repetitive training can be established as the target repetitive training. In the sixth iterative training, it is necessary to train the deep learning model using 64 samples in one sample data set. If the quantity of sample data that has already been trained is 30 and the duration of the training iteration is currently 4 minutes, the data throughput is, for example, 30/4.

A3 is that the data throughput when training the deep learning model based on the sample data is less than the second threshold value.

Here, the second threshold value is less than the first threshold value.

If the data throughput during execution training of the main process for the deep learning model is less than the second threshold, it is considered too slow to consume the sample data stored in the prefetch sample data queue. Here, if a larger prefetch quantity is maintained, the sample data stored in the prefetch sample data queue is always accumulated and may occupy a larger memory. Therefore, it is conceivable to reduce the quantity of sample data prefetched in the prefetch sample data queue and trigger the first update process for the prefetch quantity of the sample data.

Here, the method of determining the data throughput is similar to the above-mentioned a2, and will not be described repeatedly.

After satisfying the prefetch quantity update condition, for example, the first update process can be executed for the prefetch quantity of the sample data by adopting the following method, that is,
Based on the total memory space currently occupied by the prefetch sample data queue and the upper limit threshold for memory usage, the first update process is executed for the prefetch quantity of the sample data to obtain the target prefetch quantity.

Illustratively, for example, it is detected whether or not the total memory space occupied by the prefetch sample data queue has reached the memory usage upper limit threshold value, and the first update process is performed for the prefetch quantity of the sample data based on the detection result. Can be executed to obtain the target prefetch quantity.

Here, the total memory space occupied by the prefetch sample data queue represents the total memory space occupied by the sample data queue corresponding to all the main processes.

Specifically, if the total memory space currently occupied by the prefetch sample data queue has not reached the upper memory usage threshold, the prefetch quantity is increased by the first adjustment step size to obtain the target prefetch quantity, and / or,
When the total memory space currently occupied by the prefetch sample data queue reaches the upper memory usage threshold, the prefetch quantity is reduced by the second adjustment step size to obtain the target prefetch quantity.

Here, the first adjustment step size represents the adjustment step size when the prefetch quantity of the sample data is increased, and the second adjustment step size represents the adjustment step size when the prefetch quantity of the sample data is decreased. ..

Here, the first adjustment step size and the second adjustment step size may have the same size or different sizes.

Illustratively, the first adjustment step size is larger than, for example, the second adjustment step size, and in such cases, when the prefetch quantity needs to be increased, the prefetch sample is increased by increasing the prefetch quantity faster. It guarantees that the sample data stored in the data queue reaches a larger quantity faster, can meet the requirements for use in subsequent training iteration cycles, and the prefetch quantity is too small to delay the model training process. It can be avoided. At the same time, when the prefetch quantity needs to be reduced, the prefetch quantity can be reduced more slowly to ensure smoother changes in the length of the prefetch sample data queue, and the quantity of prefetch sample data. It is possible to avoid the shock of the training process due to the sudden decrease of.

Furthermore, in another embodiment, the prefetch quantity of the sample data is further based on the total memory space currently occupied by the prefetch sample data queue, the upper limit of memory usage, and the data throughput for training the deep learning model. The target prefetch quantity can be obtained by executing the first update process for the data.

Here, based on the above embodiment, for example, if the total memory space currently occupied by the prefetch sample data queue has not reached the memory usage upper threshold.
If the data throughput for training the deep learning model satisfies a predetermined data throughput condition, the prefetch quantity is increased by the first adjustment step size to obtain the target prefetch quantity.

In another embodiment, if the total memory space currently occupied by the prefetch sample data queue does not reach the memory usage upper limit threshold and the data throughput not yet meets the predetermined data throughput condition, the prefetch quantity is adjusted to the third adjustment. It further includes reducing the step size by the step size to get the target prefetch quantity.

Specifically, the first adjustment step size and the third adjustment step size may have the same size or different sizes. Similarly, the second adjustment step size and the third adjustment step size may have the same size or different sizes.

In another embodiment, if the total memory space currently occupied by the prefetch sample data queue has reached the upper memory usage threshold, whether the data throughput for training the deep learning model meets a predetermined data throughput condition. Regardless, the prefetch quantity is reduced by the second adjustment step size to get the target prefetch quantity.

The above predetermined data throughput condition includes at least one of the following b1 to b2.
b1 is that the current value of the data throughput is larger than the historical value, where the historical value is the average value of the data throughput corresponding to the plurality of historical iterative trainings before the current iterative training, or the current value. It is a numerical value of the data throughput of one iterative training before the iterative training.

For a specific determination method, for example, a2.1 above may be referred to, and the description will not be repeated here.

b2 is that the current value of the data throughput is larger than the data throughput threshold.

Here, the current numerical value of the data throughput may be referred to, for example, a2.2 above, and will not be repeatedly described here.

Further, the data processing method provided by the embodiment of the present invention is based on the above embodiment.
It further includes performing an adjustment step size second update process for the prefetch quantity to obtain the target adjustment step size, where the target adjustment step size is used for the next one update process of the prefetch quantity.

Here, for example, when the prefetch quantity is increased in the first update process, the adjustment step size of the prefetch quantity is increased to obtain the target adjustment step size, and / or.
In the case of the reduced prefetch quantity during the first update process, the target adjustment step size after the adjustment step size of the reduced prefetch quantity is obtained.

Specific examples are as follows.

A total of five processes, such as M1, M2, M3, M4, and M5, are executing training tasks of the same deep learning model in parallel.

Here, M1, M2, M3, M4, and M5 each execute the data processing method provided by the embodiment of the present invention.

Take an example in which M1 executes the data processing method.

Example 1: The first update process is executed for the prefetch quantity of the sample data based on the memory usage upper limit threshold.

1.1 By M1, the prefetch sample data queue L1 corresponding to M1, the prefetch sample data queue L2 corresponding to M2, the prefetch sample data queue L3 corresponding to M3, the prefetch sample data queue L4 corresponding to M4, and M5 Detects whether or not the total memory space occupied by the prefetch sample data queue L5 corresponding to the above has reached the upper limit of memory usage, and if not, 1.2 (a) and 1.2 (b). ), And if the application fails when M1 applies for memory to the main process of the operation system, the transition to 1.3 occurs.

1.2 (a): By M1, the first update process is executed for the prefetch quantity of the sample data to obtain the target prefetch quantity.
"Target prefetch quantity = prefetch quantity + first adjustment step size", where the first adjustment step size is the target adjustment obtained by executing the second update process for the previous one adjustment step size. The step size.

1.2 (b): The second update process is executed for the first adjustment step size by M1.
"Target adjustment step size obtained after this second update process = first adjustment step size * 2", that is, the first adjustment step size used for the next one first update process is this time. It is twice the adjustment step size used for the first update process of.

1.3: M1 detects whether the second adjustment step size is larger than 1, and if the second adjustment step size is larger than 1, it becomes 1.4 (a) and 1.4 (b). It transitions, and if it is not large, it transitions to 1.5.

1.4 (a): The second update process is executed for the second adjustment step size by M1.
"Second adjustment step size after adjustment = second adjustment step size before adjustment / 2".

1.4 (b): By M1, the first update process is executed for the prefetch quantity of the sample data to obtain the target prefetch quantity.
"Target prefetch quantity = prefetch quantity-second adjustment step size". Here, the second adjustment step size is the adjusted adjustment step size in 1.4 (a).

1.5: By M1, the second adjustment step size is maintained as it is, and based on the maintained second adjustment step size, the first update process is executed for the prefetch quantity of the sample data to set the target prefetch quantity. Get,
"Target prefetch quantity = prefetch quantity-second adjustment step size maintained as it is".

Example 2: According to M1, the first update process is executed for the prefetch quantity of the sample data based on the memory usage upper limit threshold value and the data throughput for training the deep learning model.

2.1 According to M1, the prefetch sample data queue L1 corresponding to M1, the prefetch sample data queue L2 corresponding to M2, the prefetch sample data queue L3 corresponding to M3, the prefetch sample data queue L4 corresponding to M4, and M5 Detects whether or not the total memory space occupied by the prefetch sample data queue L5 corresponding to the above has reached the memory usage upper limit threshold, and if not, transitions to 2.2, and if it is reached or M1. If the application fails when applying for memory to the main process of the operation system, the transition to 2.7 occurs.

2.2: M1 detects whether or not the data throughput when training a deep learning model satisfies a predetermined data throughput condition, and if it is satisfied, it transitions to 2.3 (a) and 2.3 (b). However, if it is not satisfied, it transitions to 2.4 (a) and 2.4 (b).

2.3 (a): By M1, the first update process is executed for the prefetch quantity of the sample data to obtain the target prefetch quantity.
"Target prefetch quantity = prefetch quantity + first adjustment step size", where the first adjustment step size is the target adjustment obtained by executing the second update process for the previous one adjustment step size. The step size.

2.3 (b): The second update process is executed for the first adjustment step size by M1.
"Target adjustment step size obtained after this second update process = first adjustment step size * 2".

2.4 (a): By M1, the first update process is executed for the prefetch quantity of the sample data to obtain the target prefetch quantity.
"Target prefetch quantity = prefetch quantity-third adjustment step size".

2.4 (b): M1 detects whether or not the third adjustment step size is larger than 1, and if the third adjustment step size is larger than 1, it transitions to 2.5, and if it is not large, it transitions to 2.5. Transition to 2.6.

2.5: By M1, the second update process is executed for the third adjustment step size.
"Third adjustment step size after adjustment = third adjustment step size before adjustment / 2".

2.6: By M1, the third adjustment step size is maintained as it is. The third adjustment step size is used when the first update process is executed for the next one prefetch quantity.

2.7: M1 detects whether the second adjustment step size is larger than 1, and if the second adjustment step size is larger than 1, it becomes 2.8 (a) and 2.8 (b). It transitions, and if it is not large, it transitions to 2.9.

2.8 (a): The second update process is executed for the second adjustment step size by M1.
"Second adjustment step size after adjustment = second adjustment step size before adjustment / 2".

2.8 (b): By M1, the first update process is executed for the prefetch quantity of the sample data to obtain the target prefetch quantity.
"Target prefetch quantity = prefetch quantity-second adjustment step size". Here, the second adjustment step size is the adjusted adjustment step size in 2.8 (a).

2.9: By M1, the second adjustment step size is maintained as it is, and based on the maintained second adjustment step size, the first update process is executed for the prefetch quantity of the sample data to set the target prefetch quantity. Get,
"Target prefetch quantity = prefetch quantity-second adjustment step size maintained as it is".

Through each step in the above example, the first update process is executed for the prefetch quantity of the sample data.

II: In S102 above, the main process may read new sample data directly from the sample database when the quantity of sample data currently contained in the prefetch sample data queue has not reached the target prefetch quantity. , You may read new sample data from the sample database by communicating with the subprocess and controlling the subprocess.

If the main process reads new sample data directly from the sample database, the main process prefetches based on the quantity of sample data extracted from the prefetch sample data queue and the quantity of sample data in the read prefetch sample data queue. Determine the quantity of sample data currently stored in the sample data queue, then compare the quantity with the target prefetch quantity, and if the quantity is less than the target prefetch quantity, pull the new sample data directly from the sample database. It can be read and stored in the prefetch sample data queue.

If the main process controls the subprocess to read new sample data from the sample database, the main process communicates with the subprocess to determine the quantity of sample data currently stored in the prefetch sample data queue. , Compare the quantity with the target prefetch quantity, and if the quantity is less than the target prefetch quantity, send a sample data read instruction to the subprocess, where the instruction to read the sample data is a sample that needs to be read. Contains quantity information for the data. After receiving the sample data read instruction sent by the main process, the subprocess can read new sample data and store it in the prefetch sample data queue based on the quantity information contained in the sample data read instruction. ..

In the embodiment of the present invention, the main process executes the first update process on the prefetch quantity to obtain the target prefetch quantity, and when the sample data amount currently contained in the data queue has not reached the target prefetch quantity. In addition, to read new sample data from the sample data pool, the main process will have already completed the reading of the sample data required for the next one iterative training after performing one iterative training. .. In most cases in practice, the time it takes for the main process to read the data is often shorter than the time it takes to perform one iterative training, so a sufficient amount of sample data is always stored in the data queue. Can be ensured to meet the use of several subsequent iterative trainings, and the sample quantity will not be read in time even if the main process takes too long to read certain sample data. It was possible to avoid delays in repetitive training and improve training efficiency.

Those skilled in the art will appreciate that, in the manner described in the specific embodiments, the writing order of each step does not constitute a restriction on the implementation process, meaning a strict execution order, but the specific execution order of each step. It should be understood that it is determined by its function and possible internal logic.

Based on the same concept of the invention, the embodiments of the present invention further provide a data processing apparatus corresponding to the data processing method, and the principle of the problem to be solved by the apparatus in the embodiments of the present invention is the embodiment of the present invention. Since it is the same as the above-mentioned data processing method in the above, the implementation of the device can refer to the implementation of the method, and the repeated part is not described repeatedly.

FIG. 2 is a schematic diagram of a data processing apparatus provided by an embodiment of the present invention, wherein the apparatus is applied to training a deep learning model including one or more processes, and the apparatus is the first. The update module 21 and the read module 22 are provided here.
The first update module 21 executes the first update process on the prefetch quantity of the sample data for one target process in the one or more processes to obtain the target prefetch quantity.
The read module 22 reads and reads new sample data in response to the fact that the quantity of sample data currently contained in the prefetch sample data queue corresponding to the target process has not reached the target prefetch quantity. The new sample data is stored in the prefetch sample data queue.

In the embodiment of the present invention, the main process performs the first update process on the prefetch quantity to obtain the target prefetch quantity, and the sample data amount currently contained in the prefetch sample data queue reaches the target prefetch quantity. In order to read new sample data from the sample data pool when it is not available, the main process should have performed one iterative training and then already read the sample data required for the next one iterative training. become. In most cases in practice, the time it takes for the main process to read the data is often shorter than the time it takes to perform one iterative training, so a sufficient amount of sample data is always stored in the data queue. Can be ensured to meet the use of several subsequent iterative trainings, and the sample quantity will not be read in time even if the main process takes too long to read certain sample data. It was possible to avoid delays in repetitive training and improve training efficiency.

In one possible embodiment, when the first update module 21 executes the first update process on the prefetch quantity of the sample data to obtain the target prefetch quantity,
Based on the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes and the upper memory usage threshold, the target prefetch is performed by performing the first update process on the prefetch quantity of the sample data. Get the quantity.

In one possible embodiment, the first update module 21 is based on the total memory space and memory usage limit threshold currently occupied by the prefetch sample data queue corresponding to the one or more processes. When the first update process is executed for the prefetch quantity to obtain the target prefetch quantity,
The total memory space currently occupied by the prefetch sample data queue for the one or more processes, the upper memory usage threshold, and the data throughput to perform training on the deep learning model of the target process. The first update process is executed for the prefetch quantity of the sample data based on the above, and the target prefetch quantity is obtained.

In one possible embodiment, the first update module 21 is based on the total memory space and memory usage limit threshold currently occupied by the prefetch sample data queue corresponding to the one or more processes. When the first update process is executed for the prefetch quantity to obtain the target prefetch quantity,
If the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes does not reach the memory usage upper threshold, the prefetch quantity is increased by the first adjustment step size to the target. Get the prefetch quantity and / or
When the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes reaches the upper memory usage threshold, the prefetch quantity is reduced by the second adjustment step size to achieve the target prefetch. Get the quantity.

In one possible embodiment, if the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes does not reach the memory usage upper limit threshold. When the prefetch quantity is increased by the first adjustment step size to obtain the target prefetch quantity,
The total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes has not reached the upper memory usage threshold and training is performed on the deep learning model of the target process. When the data throughput for this is satisfied with a predetermined data throughput condition, the prefetch quantity is increased by the first adjustment step size to obtain the target prefetch quantity.

In one possible embodiment, the first update module 21 further
When the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes does not reach the memory usage upper limit threshold, and the data throughput does not meet the predetermined data throughput condition. , The prefetch quantity is reduced by the third adjustment step size to obtain the target prefetch quantity.

In one possible embodiment, the predetermined data throughput condition is
The current value of the data throughput is larger than the historical value, and
The current value of the data throughput is greater than the data throughput threshold and includes at least one of, where the historical value is the data from a plurality of historical iterative training prior to the current iterative training. It is the average value of the throughput, or the numerical value of the data throughput at the time of one repetitive training before the current repetitive training.

In one possible embodiment, the apparatus further comprises a second update module 23 for performing a second update process on the prefetch quantity adjustment step size to obtain a target adjustment step size, wherein said. The target adjustment step size is used for the next one-time update process of the prefetch quantity.

In one possible embodiment, the second update module 23 performs a second update process on the adjustment step size of the prefetch quantity to obtain a target adjustment step size.
When the prefetch quantity is increased during the first update process, the adjustment step size of the prefetch quantity is increased and / or.
When the prefetch quantity is reduced during the first update process, the adjustment step size of the prefetch quantity is reduced.

The description of the processing flow of each module of the apparatus and the interaction flow between each module can refer to the relevant description of the embodiments of the above method and will not be repeated herein.

The embodiments of the present invention further provide the computer device 30. FIG. 3 is a schematic view showing the configuration of the computer device 30 provided by the embodiment of the present invention.
It includes a processor 31, a memory 32, and a bus 33. The memory 32 stores instructions, and the memory 32 includes an internal memory 321 and an external memory 322. The internal memory 321 here is also referred to as a memory, and includes arithmetic data of the processor 31 and an external memory 322 such as a hard disk. Temporarily store the data to be exchanged. The processor 31 executes data exchange with the external memory 322 via the internal memory 321, and when the computer device 300 is operated, the processor 31 and the memory 32 communicate with each other via the bus 33. , The processor 31 executes the following instruction, and the instruction is
For one target process in the one or more processes, the first update process is executed for the prefetch quantity of the sample data to obtain the target prefetch quantity.
The new sample data is read and read in response to the fact that the quantity of sample data currently contained in the prefetch sample data queue corresponding to the target process has not reached the target prefetch quantity. To store the prefetch sample data queue.

In one possible embodiment, in the instruction executed by the processor 31, the first update process is executed for the prefetch quantity of the sample data to obtain the target prefetch quantity.
Based on the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes and the upper memory usage threshold, the target prefetch is performed by performing the first update process on the prefetch quantity of the sample data. Includes getting quantity.

In one possible embodiment, in an instruction executed by processor 31, based on the total memory space and memory usage upper bound threshold currently occupied by the prefetch sample data queue corresponding to the one or more processes. To obtain the target prefetch quantity by executing the first update process for the prefetch quantity of the sample data,
The total memory space currently occupied by the prefetch sample data queue for the one or more processes, the upper memory usage threshold, and the data throughput to perform training on the deep learning model of the target process. The first update process is executed for the prefetch quantity of the sample data based on the above, and the target prefetch quantity is obtained.

In one possible embodiment, in an instruction executed by processor 31, based on the total memory space and memory usage upper bound threshold currently occupied by the prefetch sample data queue corresponding to the one or more processes. To obtain the target prefetch quantity by executing the first update process for the prefetch quantity of the sample data,
If the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes does not reach the memory usage upper threshold, the prefetch quantity is increased by the first adjustment step size to the target. Get the prefetch quantity and / or
When the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes reaches the upper memory usage threshold, the prefetch quantity is reduced by the second adjustment step size to achieve the target prefetch. Includes getting quantity.

In one possible embodiment, among the instructions executed by the processor 31, the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes reaches the memory usage upper limit threshold. If not, increasing the prefetch quantity by the first adjustment step size to obtain the target prefetch quantity
The total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes has not reached the upper memory usage threshold and training is performed on the deep learning model of the target process. When the data throughput for the purpose satisfies a predetermined data throughput condition, the prefetch quantity is increased by the first adjustment step size to obtain the target prefetch quantity.

In one possible embodiment, among the instructions executed by the processor 31, the method is:
When the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes does not reach the memory usage upper limit threshold, and the data throughput does not meet the predetermined data throughput condition. Further includes reducing the prefetch quantity by a third adjustment step size to obtain the target prefetch quantity.

In one possible embodiment, in the instruction executed by the processor 31, the predetermined data throughput condition is
The current value of the data throughput is larger than the historical value, and
The current value of the data throughput is greater than the data throughput threshold and includes at least one of, where the historical value is the data from a plurality of historical iterative training prior to the current iterative training. It is the average value of the throughput, or the numerical value of the data throughput at the time of one repetitive training before the current repetitive training.

In one possible embodiment, among the instructions executed by the processor 31, the method is:
Further including executing a second update process for the adjustment step size of the prefetch quantity to obtain a target adjustment step size, where the target adjustment step size is the next one update process of the prefetch quantity. Used for.

In one possible embodiment, in the instruction executed by the processor 31, the second update process is executed for the adjustment step size of the prefetch quantity to obtain the target adjustment step size.
When the prefetch quantity is increased during the first update process, the adjustment step size of the prefetch quantity is increased and / or.
When the prefetch quantity is reduced during the first update process, it includes reducing the adjustment step size of the prefetch quantity.

Embodiments of the present invention further provide a computer-readable recording medium, wherein a computer program is recorded in the computer-readable recording medium, and when the computer program is operated by a processor, an embodiment of the above-mentioned method is performed. The steps of the data processing method described in are performed. Here, the recording medium can be a volatile or non-volatile computer-readable recording medium.

The computer program product of the data processing method provided by the embodiment of the present invention includes a computer-readable recording medium containing the program code, and the instructions contained in the program code are described in the above-described embodiment of the method. The steps of the data processing method can be performed, specifically, examples of the above-mentioned method can be referred to, which are not repeated here.

The embodiments of the present invention further provide a computer program, and when the computer program is executed by a processor, any method of the above-described embodiment is realized. The computer program product can be specifically realized by a method of hardware, software, or a combination thereof. In one example of the option, the computer program product is specifically embodied by a computer recording medium. In another embodiment of the option, the computer program product is specifically embodied by a software product such as a Software Development Kit (SDK).

For convenience and simplification of description, those skilled in the art can refer to the corresponding processes of the embodiments of the above-mentioned methods for the specific working processes of the systems and devices described above, which are repeated herein. It should be clearly understood that it will not be explained. It should be understood that in some of the embodiments provided by the present invention, the disclosed systems, devices, and methods may be implemented by other methods. The embodiments of the device described above are merely exemplary. For example, the division of the unit is only a logical function division, and there may be other division methods in the actual implementation. Also, for example, multiple units or components may be combined, integrated into another system, or some features may be ignored or not performed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed interactions can be via an indirect coupling or communication connection of several communication interfaces, devices or units, in electrical, mechanical or other forms. possible.

A unit described as a separated part may or may not be physically separated, and may or may not be a physical part physical unit displayed as a unit. That is, it can be located in one location or distributed across multiple network units. The objectives of the solution of this embodiment may be achieved by selecting some or all of the units here as needed in practice.

Also, each functional unit in each embodiment of the invention may be integrated into one processing unit, or each unit may physically exist independently, or two or more units may be one. Can be integrated into a unit.

When the function is implemented in the form of a software functional unit and sold or used as a stand-alone product, it can be stored in a non-volatile computer-readable storage medium run by the processor. Based on this understanding, the essence of the technical solution of the invention, or any part of the prior art or part of the technical solution, can be embodied in the form of a software product. The computer software product is housed in one recording medium and the computer device (such as a personal computer, server, or network device) can perform all or part of the steps of the methods described in each embodiment of the invention. Includes some instructions to make it so. The recording media described above include U disks, mobile hard disks, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disks, optical disks, and various other recording media that can store program codes. Can include various media.

Finally, the above embodiments are merely specific embodiments of the invention and are used to illustrate the technical solutions of the invention rather than limiting them. The scope of protection is not limited to this. Although the invention has been described in detail with reference to the embodiments described above, those skilled in the art will modify or modify the technical solutions described in the embodiments described above within the technical scope disclosed in the present invention. It should be understood that changes can be easily conceived or some technical features here can be replaced equally. These modifications, changes, or replacements do not cause the essence of the applicable technical solution to deviate from the spirit and scope of the technical solution of the embodiments of the invention, and are all within the scope of the invention. Should be covered. Therefore, the scope of protection of the present invention should be in accordance with the scope of claims.

Claims (21)

A data processing method that applies to training deep learning models, including one or more processes.
For one target process in the one or more processes, the first update process is executed for the prefetch quantity of the sample data to obtain the target prefetch quantity.
The new sample data is read and read in response to the fact that the quantity of sample data currently contained in the prefetch sample data queue corresponding to the target process has not reached the target prefetch quantity. A data processing method comprising storing the prefetch sample data queue. To obtain the target prefetch quantity by executing the first update process for the prefetch quantity of the sample data.
Based on the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes and the upper limit memory usage threshold, the target is to execute the first update process for the prefetch quantity of the sample data. The data processing method according to claim 1, wherein the prefetch quantity is obtained. Based on the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes and the memory usage upper limit threshold, the target is to execute the first update process for the prefetch quantity of the sample data. Getting the prefetch quantity is
The total memory space currently occupied by the prefetch sample data queue for the one or more processes, the upper memory usage threshold, and the data throughput to perform training on the deep learning model of the target process. The data processing method according to claim 2, further comprising performing a first update process on the prefetch quantity of the sample data to obtain the target prefetch quantity. Based on the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes and the memory usage upper limit threshold, the target is to execute the first update process for the prefetch quantity of the sample data. Getting the prefetch quantity is
If the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes does not reach the memory usage upper threshold, the prefetch quantity is increased by the first adjustment step size to the target. Get the prefetch quantity and / or
When the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes reaches the memory usage upper limit threshold, the prefetch quantity is reduced by the second adjustment step size to achieve the target prefetch. The data processing method according to claim 2 or 3, wherein the data processing method comprises obtaining a quantity. If the total memory space currently occupied by the prefetch sample data queue for the one or more processes does not reach the memory usage upper threshold, the prefetch quantity is increased by the first adjustment step size to the target. Getting the prefetch quantity is
The total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes has not reached the upper memory usage threshold and training is performed on the deep learning model of the target process. The data processing according to claim 4, wherein when the data throughput for performing the data throughput satisfies a predetermined data throughput condition, the prefetch quantity is increased by the first adjustment step size to obtain the target prefetch quantity. Method. When the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes does not reach the memory usage upper limit threshold, and the data throughput does not meet the predetermined data throughput condition. The data processing method according to claim 5, further comprising reducing the prefetch quantity by a third adjustment step size to obtain the target prefetch quantity. The predetermined data throughput condition is
The current value of the data throughput is larger than the historical value, and
The current value of the data throughput is greater than the data throughput threshold and includes at least one of them.
Here, the historical numerical value is the average value of the data throughput at the time of a plurality of historical iterative trainings before the current iterative training, or the data throughput at the time of one iterative training before the current iterative training. The data processing method according to claim 5 or 6, wherein the data is a numerical value of. Further including executing a second update process for the adjustment step size of the prefetch quantity to obtain the target adjustment step size.
Here, the data processing method according to any one of claims 1 to 7, wherein the target adjustment step size is used for the next one-time update processing of the prefetch quantity. To obtain the target adjustment step size by executing the second update process for the adjustment step size of the prefetch quantity.
When the prefetch quantity is increased during the first update process, the adjustment step size of the prefetch quantity is increased and / or.
The data processing method according to claim 8, wherein when the prefetch quantity is reduced during the first update process, the adjustment step size of the prefetch quantity is reduced. A data processing device that is applied to the training of deep learning models, including one or more processes.
For one target process in the one or more processes, a first update module for executing a first update process for a prefetch quantity of sample data to obtain a target prefetch quantity, and
The new sample data is read and read in response to the fact that the quantity of sample data currently contained in the prefetch sample data queue corresponding to the target process has not reached the target prefetch quantity. A data processing device comprising a read module for storing the prefetch sample data queue. When the first update module executes the first update process on the prefetch quantity of the sample data to obtain the target prefetch quantity,
Based on the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes and the upper limit memory usage threshold, the target is to execute the first update process for the prefetch quantity of the sample data. The data processing apparatus according to claim 10, wherein a prefetch quantity is obtained. The first update module is first with respect to the prefetch quantity of the sample data based on the total memory space and memory usage upper bound threshold currently occupied by the prefetch sample data queue corresponding to the one or more processes. When performing the update process to get the target prefetch quantity
The total memory space currently occupied by the prefetch sample data queue for the one or more processes, the upper memory usage threshold, and the data throughput to perform training on the deep learning model of the target process. The data processing apparatus according to claim 11, wherein the first update process is performed on the prefetch quantity of the sample data to obtain the target prefetch quantity. The first update module is first with respect to the prefetch quantity of the sample data based on the total memory space and memory usage upper bound threshold currently occupied by the prefetch sample data queue corresponding to the one or more processes. When performing the update process to get the target prefetch quantity
If the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes does not reach the memory usage upper threshold, the prefetch quantity is increased by the first adjustment step size to the target. Get the prefetch quantity and / or
When the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes reaches the memory usage upper limit threshold, the prefetch quantity is reduced by the second adjustment step size to achieve the target prefetch. The data processing apparatus according to claim 11 or 12, wherein the quantity is obtained. The first update module adjusts the prefetch quantity to the first adjustment step when the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes does not reach the memory usage upper limit threshold. When increasing by size to get the target prefetch quantity
The total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes has not reached the upper memory usage threshold and training is performed on the deep learning model of the target process. The data processing apparatus according to claim 13, wherein the prefetch quantity is increased by the first adjustment step size to obtain the target prefetch quantity when the data throughput for performing the data throughput satisfies a predetermined data throughput condition. The first update module further
When the total memory space currently occupied by the prefetch sample data queue corresponding to the one or more processes does not reach the memory usage upper limit threshold, and the data throughput does not meet the predetermined data throughput condition. The data processing apparatus according to claim 14, wherein the prefetch quantity is reduced by a third adjustment step size to obtain the target prefetch quantity. The predetermined data throughput condition is
The current value of the data throughput is larger than the historical value, and
The current value of the data throughput is greater than the data throughput threshold and includes at least one of them.
Here, the historical numerical value is the average value of the data throughput at the time of a plurality of historical iterative trainings before the current iterative training, or the data throughput at the time of one iterative training before the current iterative training. The data processing apparatus according to claim 14 or 15, characterized in that the numerical value of the above. Further, a second update module for executing a second update process for the adjustment step size of the prefetch quantity to obtain a target adjustment step size is provided.
The data processing apparatus according to any one of claims 10 to 16, wherein the target adjustment step size is used for the next one-time update process of the prefetch quantity. When the second update module executes the second update process for the adjustment step size of the prefetch quantity to obtain the target adjustment step size,
When the prefetch quantity is increased during the first update process, the adjustment step size of the prefetch quantity is increased and / or.
The data processing apparatus according to claim 17, wherein when the prefetch quantity is reduced during the first update process, the adjustment step size of the prefetch quantity is reduced. It ’s a computer device,
Equipped with a processor, a recording medium, a bus,
Machine-readable instructions that can be executed by the processor are recorded on the recording medium.
When the computer device is operated, the processor and the recording medium communicate with each other via a bus, and the machine-readable instruction is executed by the processor according to any one of claims 1 to 9. A computer device characterized by performing the described data processing method. A computer-readable recording medium
A computer program is recorded in the computer-readable recording medium, and when the computer program is operated by a processor, the data processing method according to any one of claims 1 to 9 is executed. A computer-readable recording medium characterized by that. It ’s a computer program,
A computer program according to any one of claims 1 to 9, wherein when the computer program is executed by a processor, the data processing method according to any one of claims 1 to 9 is executed.

Images