Scheduling optimization method for parallel processing of machine head data and GPU (Graphics Processing Unit)

Abstract

The invention relates to the technical field of data scheduling, in particular to a scheduling optimization method for parallel processing of head data and a GPU, and the method comprises the steps: obtaining all data of a GPU global memory and to-be-analyzed data in a head command; according to the distribution condition of the hash key values of all the data and the real-time access condition in the cache data table, updating the arrangement priority level of the hash key values in the cache data table of the GPU shared memory; and updating the cache data table in the GPU shared memory by using the arrangement priority level of the hash key values of the data to be analyzed. According to the method, the scheduling of parallel processing of the machine head data and the GPU is optimized, the problem of low data scheduling efficiency in the prior art is avoided, and the data processing efficiency and the GPU operation efficiency can be effectively improved.

Description

Technical Field

[0001] The present application relates to the technical field of data scheduling, and in particular to a scheduling optimization method for parallel processing of head data and GPU. Background Art

[0002] Head data, as the core input of the server front end, carries a variety of requests initiated by users, and these data are input into the GPU of the server. The GPU divides the data into countless small blocks through parallel processing technology and processes them simultaneously. This parallel processing capability not only greatly improves the speed and efficiency of data processing, but also maintains the stability and response speed of the server when facing massive concurrent requests, ensuring smooth data flow. In addition, GPU parallel processing also helps to achieve load balancing, by intelligently allocating tasks to different GPU cores, improving the processing power and reliability of the entire system. When a core fails, other cores can quickly take over the task to ensure the continuity of data processing and the stability of the system.

[0003] For the parallel processing of head data and GPU, part of the data will be stored in the shared memory for data caching during the processing process, so that the GPU can quickly schedule it during multi-threaded parallel processing. Traditional technology for storing shared memory is to form a FIFO queue for data, and each time the head of the linked list is updated, the tail of the linked list is eliminated. This method easily leads to less frequent data being stored in the cache, so that the data occupies the cache space. When the GPU multi-threads schedule data, they need to traverse the data, which affects the time for multi-threaded data scheduling, and ultimately leads to problems with low data processing efficiency and GPU operation efficiency. Summary of the invention

[0004] In order to solve the above technical problems, the present application provides a scheduling optimization method for parallel processing of head data and GPU to solve the existing problems.

[0005] The scheduling optimization method of the present application for head data and GPU parallel processing adopts the following technical solution: An embodiment of the present application provides a scheduling optimization method for parallel processing of head data and GPU, the method comprising the following steps: Step 1: Get all the data in the GPU global memory and the data to be analyzed in the head command; Step 2: Update the arrangement priority of the hash key values ​​in the cache data table of the GPU shared memory according to the distribution of the hash key values ​​of all data and the real-time access in the cache data table; specifically: A1, by analyzing the number of data types and data confusion under the same hash key value, calculate the mapping conflict amount of each hash key value; A2, combining the mapping relationship conflict amount, the probability of the data with the highest frequency of each hash key value in all data, and the maximum frequency of each hash key value in all data types, calculate the arrangement priority of each hash key value; select a preset number of hash key values ​​according to the size of the arrangement priority, and store them in the cache data table of the GPU shared memory; A3, analyze the average level of the difference values ​​of all hash key values ​​in the real-time cache data table after being differentiated according to the time sequence of entering the cache data table, and the average level of the difference values ​​after being differentiated according to the arrangement priority of all hash key values, and calculate the priority update amount of the hash key values ​​in the cache data table; A4, using the priority update amount, updates the arrangement priority of the hash key values ​​in the cache data table; Step 3: Update the cache data table in the GPU shared memory using the arrangement priority of the hash key value of the data to be analyzed.

[0006] Preferably, the mapping relationship conflict amount of each hash key value is determined by the product of the number of data types that appear and the disorder of the data.

[0007] Preferably, the disorder of the data is further determined by the information entropy of all data under the same hash key value.

[0008] Preferably, the calculation method of the arrangement priority of each hash key value is: Calculate the product of the probability of each hash key value appearing most frequently in all data and the maximum frequency of each hash key value appearing in all its data types; The ratio of the product to the conflict amount of the mapping relationship is used as the arrangement priority level of each hash key value.

[0009] Preferably, the method for screening out a preset number of hash key values ​​is further determined as follows: arranging all hash key values ​​in descending order of priority, and screening out hash key values ​​corresponding to the first preset number of arrangement priorities.

[0010] Preferably, the priority update amount of the hash key value in the cache data table is determined by the product of the average levels of the two differential values ​​in step A3.

[0011] Preferably, the average level of the post-differentiation difference values ​​is further determined by the average of the absolute values ​​of all post-differentiation difference values.

[0012] Preferably, the method for updating the arrangement priority of the hash key values ​​in the cache data table is: For each hash key value in the cache data table, when the scheduled data to be analyzed is the data corresponding to the hash key value, the arrangement priority of the hash key value is added to the priority update amount; otherwise, the arrangement priority of the hash key value is subtracted from the priority update amount.

[0013] Preferably, the method for updating the cache data table in the GPU shared memory is: When the hash key value of the data to be analyzed is in the cache data table, all the hash key values ​​in the cache data table are reordered from large to small according to the updated arrangement priority; Otherwise, the arrangement priority of the hash key value of the data to be analyzed is used to determine whether the corresponding hash key value is added to the cache data table.

[0014] Preferably, the method for determining whether the corresponding hash key value is added to the cache data table is: When the arrangement priority of the hash key value of the data to be analyzed is greater than or equal to the arrangement priority of the last hash key value in the cache data table, the hash key value of the data to be analyzed replaces the last hash key value in the cache data table; Otherwise, no adjustment is made to the cache data table.

[0015] This application has at least the following beneficial effects: The traditional method uses FIFO queues to manage cache data in GPU shared memory, which causes lower frequency data to occupy cache space, affecting data scheduling efficiency. The present application optimizes the scheduling of head data and GPU parallel processing, obtains the data to be analyzed in the head data in real time, analyzes the number of data types and data confusion that appear under the same hash key value of all data, calculates the mapping relationship conflict amount of each hash key value to reflect the conflict amount between data; combines the mapping relationship conflict amount, the probability of each hash key value appearing most frequently in all data and the maximum frequency of each hash key value appearing in all its data types, calculates the arrangement priority of each hash key value to reflect the priority of data; analyzes the average level of differential values ​​of all hash key values ​​in the real-time cache data table after differentiating according to the time sequence of entering the cache data table, and the average level of differential values ​​after differentiating according to the arrangement priority of all hash key values, calculates the priority update amount of the hash key value in the cache data table, and uses the priority update amount to update the arrangement priority of the hash key value in the cache data table, that is, dynamically adjusts the arrangement priority of the hash key value in combination with the real-time data scheduling situation, and ensures that the arrangement priority of the hash key value changes continuously with the GPU running time. This application optimizes the scheduling of head data and GPU parallel processing, avoids the problem of low data scheduling efficiency in the prior art, and can effectively improve data processing efficiency and GPU operation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present application or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 A flowchart of a scheduling optimization method for head data and GPU parallel processing provided in this application; Figure 2 A process flow chart of updating the arrangement priority of hash key values ​​in the cache data table of GPU shared memory provided by this application. DETAILED DESCRIPTION

[0018] In order to further explain the technical means and effects adopted by the present application to achieve the predetermined invention purpose, the following is a detailed description of the scheduling optimization method for head data and GPU parallel processing proposed in the present application, its specific implementation method, structure, features and effects, in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0019] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0020] The following is a detailed description of a specific solution of a scheduling optimization method for head data and GPU parallel processing provided by the present application in conjunction with the accompanying drawings.

[0021] An embodiment of the present application provides a scheduling optimization method for parallel processing of head data and GPU.

[0022] Specifically, the following scheduling optimization method for head data and GPU parallel processing is provided, please refer to Figure 1 , the method comprises the following steps: Step 1: Get all the data in the GPU global memory and the data to be analyzed in the head command; Before the server's GPU runs computing tasks, the server receives various request commands from the user through the head, obtains the head data, and transmits the head data to the server's CPU. The CPU analyzes the commands in the head data and transmits the data that requires the GPU to participate in the calculation to the GPU's global memory. Global memory is the main memory space of the GPU and can be accessed by all cores.

[0023] Next, the GPU will process the data in blocks. In NVIDIA GPU, the basic computing unit is the Streaming Multiprocessor (SM), each SM can execute multiple threads in parallel. When the thread schedules data, the GPU has a shared memory to cache data for fast scheduling of threads in the SM. After the GPU calculation is completed, the result of the GPU calculation is transmitted to the CPU, which will package the head data and the CPU calculation result into response data and transmit it to the user through the network.

[0024] At this point, all the data in the GPU global memory and the data to be analyzed in the head command can be obtained.

[0025] Step 2: Update the arrangement priority of the hash key values ​​in the cache data table of the GPU shared memory according to the distribution of the hash key values ​​of all data and the real-time access in the cache data table; In this application, the process flow chart of updating the priority ranking of hash key values ​​in the cache data table of GPU shared memory is as shown in the attached figure. Figure 2 As shown, specifically: A1, by analyzing the number of data types and data confusion under the same hash key value, calculate the mapping conflict amount of each hash key value; As for the shared memory of the GPU, since the shared memory of the GPU is relatively small and can store relatively little data, a hash table is used in this application to schedule a hash function to perform a hash operation on the data in the global memory to obtain a hash key value of the data. The shared memory of the GPU stores the hash key value of the data, and when searching, a mapping search is performed through the hash key value. Among them, the hash function operation of the data and the search for data through the hash key value are well-known technologies, and the specific calculation process will not be repeated here.

[0026] The hash key value after all the data in the global memory is mapped by the hash function. Since hash mapping may cause hash conflicts, different data may be mapped to the same hash key value. Therefore, if a hash key value corresponds to multiple data, it takes more time to search for data through the hash key value, which reduces the search effect. Therefore, when arranging the hash key value in the GPU shared memory, it should be arranged behind the cache data table so that other data can be searched in advance.

[0027] Thus, the mapping conflict amount of each hash key value is calculated, and the mapping conflict amount of the i-th hash key value in the global memory is used. For example, the expression is: ; In the formula, Indicates the amount of mapping conflicts for the i-th hash key value; Indicates the number of types of data corresponding to the i-th hash key value; The information entropy of all data corresponding to the ith hash key value is represented. The calculation of information entropy is a well-known technology, and the specific calculation process is not repeated here. It is used to characterize the disorder of all data corresponding to the ith hash key value. In other implementations, variance can also be used for characterization.

[0028] For a hash key value of data in the global memory of a GPU, the more types of data the hash key value corresponds to, that is, The larger the value, the greater the possibility of conflict when mapping data to hash key values. In the process of GPU thread scheduling data, the time required for scheduling through the hash key value is longer than the hash key value corresponding to fewer data types. The more chaotic the frequency of occurrence of data corresponding to the hash key value, the more uneven the frequency of scheduling of different data corresponding to the hash key value. Therefore, the value of information entropy of the frequency of occurrence of data corresponding to the hash key value is The bigger.

[0029] Therefore, the hash key value should be arranged behind the cache data table during the arrangement process, so that data that can be queried quickly can be scheduled for query first, thereby reducing the query scheduling time of the data.

[0030] In other embodiments of the present application, the mapping relationship conflict amount of the i-th hash key value may also be Denoted as: .

[0031] A2, combining the mapping relationship conflict amount, the probability of the data with the highest frequency of each hash key value in all data, and the maximum frequency of each hash key value in all data types, calculate the arrangement priority of each hash key value; select a preset number of hash key values ​​according to the size of the arrangement priority, and store them in the cache data table of the GPU shared memory; When searching for data through hash key values, the GPU thread needs to traverse the cache data table of the shared memory first. If the frequency of data occurrence is higher, it means that the data is scheduled more times by the threads in the GPU. Arranging its hash key value in front of the cache data table can effectively reduce the time for the thread to schedule data. When the hash key value corresponds to multiple data, the data with the highest frequency of occurrence is given priority, which can reduce data conflicts during traversal.

[0032] Therefore, we can calculate the ranking priority of each hash key value by analyzing the frequency of data occurrence, and use the ranking priority of the kth hash key value in the global memory as the ranking priority. For example, the expression is: ; In the formula, Indicates the ranking priority of the kth hash key value; Indicates the probability of the kth hash key value being the most frequent data type in all data; Indicates the maximum frequency of occurrence of the kth hash key value in all its data types; represents the mapping conflict amount of the kth hash key value. The calculation of data probability is a well-known technology, and the specific calculation process is not repeated here.

[0033] For multi-threaded parallel computing on GPU, different threads may use the same data. Therefore, the number of scheduling times for different data is different. The more times the data is scheduled by GPU threads, the more times the data is used, and the higher the probability of scheduling the data in the future. For the data corresponding to the hash key value, the fewer the types of data corresponding to the hash key value, the less time is required for scheduling the data. Therefore, the maximum probability value of the data is used. As the scheduling probability weight; therefore, the frequency of occurrence of this data As a possibility, the larger the two values ​​are, the higher the possibility of future scheduling of the data corresponding to the hash key value is. At the same time, the mapping conflict of the hash key value is The smaller it is, the less time it takes to schedule the data corresponding to the hash key value, which increases the priority of the hash key value.

[0034] Then, all hash key values ​​are arranged in descending order according to their arrangement priority, and the first N hash key values ​​after arrangement are selected, where N is 32 in this embodiment and can be set according to actual conditions. The selected hash key values ​​are stored in the cache data table of the shared memory of the GPU.

[0035] A3, analyze the average level of the difference values ​​of all hash key values ​​in the real-time cache data table after being differentiated according to the time sequence of entering the cache data table, and the average level of the difference values ​​after being differentiated according to the arrangement priority of all hash key values, and calculate the priority update amount of the hash key values ​​in the cache data table; When GPU performs multi-threaded operations, it may happen that a large number of threads need the same data in the initial state, causing the data to be scheduled more times, thus making the hash key value corresponding to the data have a higher priority. However, as the program executes, some data may no longer be frequently accessed, but their corresponding hash key values ​​may still remain in the cache, causing the hash key value to be repeatedly traversed during the subsequent thread scheduling process, increasing the time for thread scheduling data. Therefore, when the GPU is running in real time, it is necessary to update the cache data table in the shared memory according to the data scheduling time.

[0036] For the cache data table in the shared memory of the GPU, in the best case, the order of the hash key values ​​is consistent with the scheduling time, that is, the closer to the current time, the higher the hash key value is arranged, which can save scheduling time.

[0037] Therefore, for the hash key value in the cache data table, the first-order difference calculation is performed according to the time sequence of the hash key value stored in the cache data table, and then the average of the absolute values ​​of all the difference values ​​is calculated to obtain the scheduling priority time difference, which is recorded as Q and used to characterize the degree of change in the data scheduling type. Among them, the calculation of the first-order difference method is a well-known technology, and the specific calculation process will not be repeated. Based on the above analysis, the priority level update amount V of the hash key value in the cache data table is calculated, and the expression is: ; In the formula, Indicates the priority update amount of the hash key value in the cache data table; Represents the scheduling priority time difference of the hash key value; Indicates the class difference of hash key values ​​in the cache data table, which is used to characterize the degree of data scheduling change. The class difference of hash key values ​​in the cache data table is calculated by taking the first-order difference of the arrangement priority of the hash key values ​​in the cache data table and then finding the average of the absolute values ​​of all the difference values.

[0038] It should be understood that when data is called continuously, the higher the priority of the data scheduled by the GPU thread, the more urgent the data is to be used, and the higher the priority of scheduling the data in the future. The higher the hash key value is in the cache table, the less the type of data scheduling has changed. Therefore, the scheduling priority time difference Q of the hash key value is smaller, the class difference R of the hash key value is smaller, and ultimately the priority level update amount V of the hash key value is smaller. When there are major changes in data scheduling, the hash key value scheduling time in the cache data table will be disordered. Therefore, the arrangement priority level of the hash key value of the currently scheduled data needs to be increased, that is, the scheduling priority time difference of the hash key value The larger the value, the higher the priority update of the hash key value in the cache data table. The value of is larger, thereby improving the efficiency of GPU multi-threaded data scheduling.

[0039] In other embodiments of the present application, the priority update amount V of the hash key value in the cache data table can also be set to .

[0040] A4, using the priority update amount, updates the arrangement priority of the hash key values ​​in the cache data table; Based on the above steps, the priority of the hash key value is updated through the real-time scheduling data to cache the updated priority of the xth hash key value in the data table. For example, the update method is: ; In the formula, Indicates the priority of the updated x-th hash key value in the cache data table; Indicates the priority of the xth hash key value in the cache data table before it is updated; Represents the scheduling label, c=1 means that the scheduled data to be analyzed is the data corresponding to the hash key value, and c=0 means that the scheduled data to be analyzed is not the data corresponding to the hash key value.

[0041] In GPU multi-threaded computing, the frequent scheduling of data has a direct impact on the arrangement priority of hash key values. When a piece of data is frequently scheduled recently, the arrangement priority of its corresponding hash key value should be increased, that is, , so that it can maintain its high priority status in the cache for fast access. On the contrary, for those data that are no longer frequently scheduled, the priority of their hash key values ​​should be reduced, that is, , so that the sorting of these data in the cache data table of the GPU shared memory gradually decreases.

[0042] This dynamic adjustment helps optimize cache usage and reduce caching of data that is no longer needed, thereby making room for more frequently accessed data. As data scheduling patterns change, the GPU needs to update the cache data table in shared memory in real time to reflect the latest data access patterns. This update can be achieved by monitoring the scheduling time of the data, thereby adjusting the arrangement priority of the hash key value. When data is frequently scheduled, increasing the priority of its hash key value can reduce the time it takes for future GPU threads to schedule data through shared memory and improve the operating efficiency of the GPU.

[0043] Step 3: Update the cache data table in the GPU shared memory using the arrangement priority of the hash key value of the data to be analyzed.

[0044] The cache data table in the GPU shared memory is updated by the above-mentioned ranking priority after the hash key value is updated at any time, specifically: When the hash key value of the data to be analyzed is in the cache data table, all the hash key values ​​in the cache data table are reordered from large to small according to the updated arrangement priority; Otherwise, the arrangement priority of the hash key value of the data to be analyzed is used to determine whether the corresponding hash key value is added to the cache data table.

[0045] The method for determining whether the corresponding hash key value is added to the cache data table is as follows: When the arrangement priority of the hash key value of the data to be analyzed is greater than or equal to the arrangement priority of the last hash key value in the cache data table, the hash key value of the data to be analyzed replaces the last hash key value in the cache data table; Otherwise, no adjustment is made to the cache data table.

[0046] This dynamic adjustment helps reduce conflicts in the cache and improves the efficiency of data retrieval. By adjusting the priority of hash key values ​​in real time, the performance degradation caused by conflicts can be reduced, ensuring that the GPU can quickly respond to data access requests, reducing the time for threads to schedule data in parallel, and improving the operating efficiency of the GPU. Each time the cache data table is updated, the multithreading in the GPU will use this updated cache data table to schedule data, and will retrieve and use data according to the latest priority order, thereby improving data retrieval efficiency.

[0047] The various embodiments in the present application are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0048] It should be noted that, unless otherwise specified and limited, terms such as "include", "comprises" or any other variants thereof are intended to cover non-exclusive inclusion, so that a circuit structure, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such article or device. In the absence of further restrictions, an element defined by the sentence "including one..." does not exclude the existence of other identical elements in the article or device including the element. In addition, the term "and\or" used herein includes any and all combinations of one or more related listed items.

[0049] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention herein. The present application is intended to cover any variations, uses or adaptations of the present application, which follow the general principles of the present application and include common knowledge or customary technical means in the art that are not invented by the present application.

[0050] It should be understood that the present application is not limited to the exact construction that has been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof.

Claims

1. A scheduling optimization method for head data and GPU parallel processing, characterized in that: The method comprises the following steps: Step 1: Get all the data in the GPU global memory and the data to be analyzed in the head command; Step 2: Update the arrangement priority of the hash key values ​​in the cache data table of the GPU shared memory according to the distribution of the hash key values ​​of all data and the real-time access in the cache data table; specifically: A1, by analyzing the number of data types and data confusion under the same hash key value, calculate the mapping conflict amount of each hash key value; A2, combining the mapping relationship conflict amount, the probability of the data with the highest frequency of each hash key value in all data, and the maximum frequency of each hash key value in all data types, calculate the arrangement priority of each hash key value; select a preset number of hash key values ​​according to the size of the arrangement priority, and store them in the cache data table of the GPU shared memory; A3, analyze the average level of the difference values ​​of all hash key values ​​in the real-time cache data table after being differentiated according to the time sequence of entering the cache data table, and the average level of the difference values ​​after being differentiated according to the arrangement priority of all hash key values, and calculate the priority update amount of the hash key values ​​in the cache data table; A4, using the priority update amount, updates the arrangement priority of the hash key values ​​in the cache data table; Step 3: Update the cache data table in the GPU shared memory using the arrangement priority of the hash key value of the data to be analyzed.

2. The scheduling optimization method for head data and GPU parallel processing according to claim 1, characterized in that: The mapping relationship conflict amount of each hash key value is determined by the product of the number of data types that appear and the disorder of the data.

3. The scheduling optimization method for head data and GPU parallel processing as claimed in claim 1, characterized in that: The disorder of the data is further determined by the information entropy of all data under the same hash key value.

4. The scheduling optimization method for head data and GPU parallel processing as claimed in claim 1, characterized in that: The calculation method of the arrangement priority of each hash key value is: Calculate the product of the probability of each hash key value appearing most frequently in all data and the maximum frequency of each hash key value appearing in all its data types; The ratio of the product to the conflict amount of the mapping relationship is used as the arrangement priority level of each hash key value.

5. The scheduling optimization method for head data and GPU parallel processing as claimed in claim 1, characterized in that: The method for screening out a preset number of hash key values ​​is further determined as follows: arranging all hash key values ​​in descending order of priority, and screening out hash key values ​​corresponding to the first preset number of arrangement priorities.

6. The scheduling optimization method for head data and GPU parallel processing as claimed in claim 1, characterized in that: The priority update amount of the hash key value in the cache data table is determined by the product of the average levels of the two differential values ​​in step A3.

7. The scheduling optimization method for head data and GPU parallel processing as claimed in claim 6, characterized in that: The average level of the post-differentiation difference values ​​is further determined by the average of the absolute values ​​of all the post-differentiation difference values.

8. The scheduling optimization method for head data and GPU parallel processing as claimed in claim 1, characterized in that: The update method of the arrangement priority of the hash key values ​​in the cache data table is: For each hash key value in the cache data table, when the scheduled data to be analyzed is the data corresponding to the hash key value, the arrangement priority of the hash key value is added to the priority update amount; otherwise, the arrangement priority of the hash key value is subtracted from the priority update amount.

9. The scheduling optimization method for head data and GPU parallel processing as claimed in claim 1, characterized in that: The update method of the cache data table in GPU shared memory is: When the hash key value of the data to be analyzed is in the cache data table, all the hash key values ​​in the cache data table are reordered from large to small according to the updated arrangement priority; Otherwise, the arrangement priority of the hash key value of the data to be analyzed is used to determine whether the corresponding hash key value is added to the cache data table.

10. The scheduling optimization method for head data and GPU parallel processing according to claim 9, characterized in that: The method for determining whether the corresponding hash key value is added to the cache data table is: When the arrangement priority of the hash key value of the data to be analyzed is greater than or equal to the arrangement priority of the last hash key value in the cache data table, the hash key value of the data to be analyzed replaces the last hash key value in the cache data table; Otherwise, no adjustment is made to the cache data table.

Images