JP2012190206A - Data processing scheduling device, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the processing efficiency of a system in which data is transferred to and processed by a co-processor by reducing waiting time for data transfer and data processing.SOLUTION: A host computer operates to divide data to be processed into multiple chunks, obtain the lengths of time needed for transferring and processing each of the chunks, determine the order in which the chunks are processed by a co-processor based on the lengths of data transfer time and data processing time of each of the chunks, transfer the chunks to the co-processor according to the determined order and direct the co-processor to process them.

Description

  The present invention relates to a data processing scheduling apparatus in a system for transferring data to a co-processor for processing.

  In recent computers, GPUs (Graphics Processing Units) have become commoditized and are widely available. The GPU is mainly used for graphics processing, but it is used as a general purpose computing on graphics processing units (GPGPU) for general computing purposes as a co-processor of the CPU (Central Processing Unit). Is attracting attention. In addition, the price of programmable devices such as field programmable gate arrays (FPGAs) that can also be used as co-processors has been reduced. These devices exchange data with a host computer via a data bus. However, since the bandwidth is limited, the data transfer time affects the overall processing time.

  In general, a data processing procedure in a system that processes a large amount of data using a coprocessor is as follows: 1) transfer of data from the host computer to the coprocessor, 2) data processing in the coprocessor, 3) co The result data is transferred from the processor to the host computer. In general, the data transfer time monotonically increases with respect to the amount of data. Data transfer time tends to be a bottleneck because the host-co-processor is connected by a bus such as PCI ExpressX16.

  In order to reduce the data transfer time and increase the overall processing speed, there are generally the following two methods. The first is a method of concealing the data transfer time by dividing the data string into a plurality of pieces and multiplexing the transfer and processing (pipeline processing). The second is a method of compressing transfer data.

  For example, in Patent Document 1, the data size and the transmission order of each compressed data to be transferred are transmitted to the device on the receiving side W, and the receiving side device transmits each compressed data to be transferred based on these pieces of information. A technique for scheduling a decompression process is disclosed.

JP 2006-338409 A

  In a system that processes data using a co-processor, the above-mentioned two methods are used to divide the data to be processed, compress the divided data (chunk), transfer it to the co-processor, and process it. Think. In this case, there may be a situation in which the data transfer time and the data processing time are different for each chunk due to different chunk sizes after compression. For example, when chunks are created with non-uniform sizes, or when chunks are created with a fixed length size, but the compression rate differs for each chunk. In this situation, when pipeline processing is performed between the host computer and the co-processor, if each chunk is transferred and processed in the original data order (in-order), data transfer wait and data There is a problem that processing wait occurs and the processing efficiency decreases.

  Even in the transfer method of Patent Document 1, since the data strings are transferred in the original order, there is a problem in that the reception side waits for data transfer and efficiency decreases.

  The present invention has been made in view of the above problems, and in a system that transfers data to a co-processor for processing, data that can reduce the waiting time for data transfer and data processing and improve processing efficiency. It is an object to provide a processing scheduling apparatus, method, and program.

  The present invention is a data processing scheduling apparatus for transferring data to a co-processor for processing, dividing means for dividing the data to be processed into a plurality of chunks, and the data transfer time and data processing time for each chunk. A time acquisition means for acquiring; and an order determination means for determining the order in which the co-processor processes the chunks based on the magnitude relationship between the data transfer time and the data processing time of each chunk. This is a data processing scheduling apparatus.

  The present invention provides a data processing scheduling method for transferring data from a host to a co-processor for processing, dividing the data to be processed into a plurality of chunks, and acquiring the data transfer time and data processing time of each chunk. The data processing scheduling method is characterized in that the order in which each of the chunks is processed by the co-processor is determined based on the magnitude relationship between the data transfer time and the data processing time of each chunk.

  The present invention provides a computer with a division process for dividing processing target data into a plurality of chunks, a time acquisition process for acquiring the data transfer time and data processing time of each chunk, and a data transfer time and a data processing time of each chunk. A program for executing an order determination process for determining an order in which the co-processor processes each chunk based on a relationship.

  According to the present invention, in a system that transfers data to a co-processor for processing, the waiting time for data transfer and data processing can be shortened and the processing efficiency can be improved.

FIG. 1 is a block diagram schematically showing a data processing scheduling apparatus according to an embodiment of the present invention. FIG. 2 is a diagram for explaining the operation of the data processing schedule device according to the present embodiment. FIG. 3 is a flowchart for explaining data registration processing. FIG. 4 is a flowchart for explaining processing order determination processing. FIG. 5 is a flowchart for explaining the execution of the process whose order has been determined. FIG. 6 is a diagram showing a specific example of the data transfer processing and data processing execution time of each chunk. FIG. 7 is a flowchart for explaining the processing when the processing result is transferred to the host side for each chunk. FIG. 8 is a flowchart for explaining the data registration process in the case where the data transfer time and the data processing time are measured and acquired.

Embodiments of the present invention will be described below with reference to the drawings. The present invention relates to a data processing schedule device for transferring data from a host computer to a co-processor for processing. Data processing to be processed by the apparatus of the present invention includes processing that is not affected by the processing order of a data string (hereinafter referred to as a chunk) obtained by dividing the data string to be processed. For example, aggregation processing such as summation calculation and average value calculation of numeric arrays, creation of bitmap index that creates a flag array with bits of specific numeric values from numeric arrays, handling millions to hundreds of millions of huge data DWH (Data
Ware House) includes internal processing.

  FIG. 1 is a block diagram schematically showing a data processing scheduling apparatus according to an embodiment of the present invention. The data processing scheduling apparatus includes a host computer 10, a co-processor 20, and a data bus 7 that connects the host computer 10 and the co-processor 20.

  The host computer 10 includes a data registration unit 1, a data storage unit 2, an order determination unit 3, and a processing instruction unit 4.

  The data registration unit 1 performs predetermined processing on the input data to be processed and registers it in the data storage unit 2. The data registration unit 1 includes a data division unit 11, a data compression unit 12, and a time acquisition unit 13.

  The data dividing unit 11 divides the input processing target data into a plurality of chunks.

  The data compression unit 12 compresses the divided chunks with a predetermined compression algorithm.

  The time acquisition unit 13 acquires the data transfer time and the data processing time based on the input data and the input processing content. The method for acquiring the data transfer time and the data processing time is arbitrary. For example, the actual measurement data based on the data size or the like may be stored in advance, and the actual measurement value may be acquired. You may acquire using the preset function.

  The data registration unit 1 corresponds to the chunks divided and compressed by the data dividing unit 11 and the data compressing unit 12, the data transfer time and the data processing time acquired by the time acquiring unit 13, and the input processing content. At the same time, it is registered in the data storage unit 2.

The data accumulation unit 2 accumulates and stores the compressed chunk, the processing content for the chunk, the data transfer time and data processing time acquired for the chuck, and the processing result for the chunk. The data storage unit 2 includes DRAM (Dynamic Random Access Memory), SSD (Solid
State Drive) etc.

  The order determination unit 3 determines the order in which each chunk is processed based on the information stored in the data storage unit 2. Details of the process for determining the order of the processes will be described later.

  The processing instruction unit 4 instructs the data transfer of each chunk and the execution of data processing according to the order determined by the order determination unit 3.

  The co-processor 20 includes a data storage unit 5 and an operation execution unit 6. The co-processor 20 includes, for example, a GPU, an FPGA, and the like.

  The data accumulation unit 5 accumulates and stores the chunk transferred from the host side, the processing for the chunk, and the processing result for the chunk. The data storage unit 5 includes a DRAM, an SSD, and the like.

  The arithmetic execution unit 6 performs data processing on the chunk that has been transferred, and stores the processing result in the data storage unit 5.

  The data bus 7 connects the host computer 10 and the co-processor 20 and transfers data from the data storage unit 2 to the data storage unit 5. The data bus 7 includes, for example, PCI Express and Ethernet (registered trademark).

  Next, the operation of the data processing scheduling apparatus according to this embodiment will be described. As shown in FIG. 2, the data processing schedule device according to the present embodiment operates in accordance with the processing flow of registration of various data (step A21), determination of processing order (step A22), and execution of processing (step A23). To do. Hereinafter, each of these steps will be described.

  First, the data registration process in step A21 will be described with reference to FIG. The data registration unit 1 receives input of processing target data, processing contents, a function for obtaining a transfer time, and a function for obtaining a processing time (step A31). The data dividing unit 11 divides the input processing target data into a plurality of chunks (step A32). The data compression unit 12 compresses each chunk (step A33). The data registration unit 1 stores each compressed chunk and the input processing content in the data storage unit 2 (step A34).

  Next, the processing order determination processing in step A22 will be described with reference to FIG. The order determination unit 2 performs the following processing for all chucks stored in the data storage unit 2 (step A41). The order determination unit 2 acquires the transfer time and the processing time for one chunk stored in the data storage unit 2 (step A42). The order determination unit 2 compares the transfer time and the processing time, and determines whether the processing time is equal to or greater than the transfer time and the magnitude relationship between them. If the transfer time is longer than the processing time (step A43: NO), the chunk is added to the list A (step A44). If the processing time is equal to or longer than the transfer time (step A43: YES), the chunk is added to the list B (step A45).

  When the processing of steps A42 to A45 is completed for all the chunks (step A41: YES), the order determining unit 2 sorts the chunks registered in the list A in the order of the shortest transfer time (step A46) and registers them in the list B. Chunks are sorted in order of processing time (step A47). The order determination unit 2 concatenates the list B after the list A, and stores it in the data storage unit 2 as the entire execution schedule (step A48).

  Thereby, the processing order of each chunk is determined. The reason why the group of chunks (list A) that is executed first and whose transfer time is shorter than the processing time is sorted in the order of short transfer time is to reduce the initial processing start waiting time. The reason why the group of chunks (list B), which are executed subsequently and whose processing time is equal to or longer than the transfer time, is sorted in descending order of processing time is to shorten the end processing time.

  Next, execution of the process of step A23 will be described with reference to FIG.

  The processing instruction unit 4 first instructs the co-processor 20 to transfer the chunk, reads the chunk from the data storage unit 2 according to the order determined by the order determination unit 3, and transmits the chunk via the data bus 7. (Step A51). The co-processor 20 stores the transferred chunk in the data storage unit 5.

  The processing instruction unit 4 reads the processing content stored in the data storage unit 2 and instructs the co-processor 20 to process the chunk (step A52). The arithmetic execution unit 6 of the co-processor 20 performs data processing instructed by the host for the chunk that has been transferred.

  The process instruction unit 4 determines whether or not the process instruction has been completed for all chunks (step A53). If it has not been completed (step A53: NO), the process instruction unit 4 returns to step A51 to perform the process for the next chunk. If the processing instructions for all the chunks have been completed (step A53: YES), the process of the last chunk by the co-processor 20 calculation execution unit 6 is awaited (step A54).

  When the processing of the last chunk is completed in the co-processor 20, the processing instruction unit 4 instructs the co-processor 20 to transfer the processing result (step A55). The co-processor 20 transfers the processing result stored in the data storage unit 5 to the host computer 10.

  Next, the operation of the data processing scheduling apparatus will be specifically described. For example, in a field such as DWH, processing of a large amount of data of millions to hundreds of millions of times may be performed many times in order to analyze data trends and relationships. Here, a case where the sum calculation of a numerical data array having 100 million elements is performed will be described as an example.

  First, a data registration process is executed (step A21). Specifically, in the data registration unit 1, a function T that returns a data transfer time with 100 million numerical data arrays, processing contents (calculating the sum of the compressed numerical strings), and a variable y as arguments. y) and a function X (y) that returns the data processing time with the variable y as an argument are input (step A31). In the present embodiment, the argument in the function T and the function X is the data size of the compressed chunk, for example, T (size) = 2α * size, X (size) = α * size.

  The data dividing unit 11 divides the input 100 million numerical data array into four 25 million chunks (chunks C0 to C3) (step A32). Since the number of chunks is 4, the co-processor 20 repeats the process of calculating the sum of the data in the chunks and adding the value to the result value stored in the data storage unit 5 four times.

  The data compression unit 12 compresses each chunk (step A33). In this example, the size of each chunk after compression is chunk [0]. size = 1, chunk [1]. size = 2, chunk [2]. size = 2, chunk [3]. Assume that size = 3.

  The data registration unit 1 stores each compressed chunk, the processing content for the compressed chunk, the function T, and the function X in the data storage unit 2 (step A34).

  Next, processing order determination processing is executed (step A22). Specifically, the order determination unit 2 determines the processing order of each chunk from the data stored in the data storage unit 2 according to the algorithm shown in FIG. The transfer time and processing time are acquired for each chunk (step A42), and the magnitude relationship between these is determined (step A43). In this example, T (size)> X (size) holds in all chunks. Therefore, each chunk is registered in the list B (step A45). When the determination of the magnitude relationship between the transfer time and the processing time is completed for all chunks (step A41: YES), the order determining unit 2 sorts the data processing time X (size) in descending order, and lists the schedule based on the sorting result. Create In this example, the chunk processing order is C3, C2, C1, and C0.

  Next, processing is executed (step A23). Specifically, the processing instruction unit 4 instructs the execution of processing in the order of chunks C3, C2, C1, and C0 according to the algorithm shown in FIG.

  The processing instruction unit 4 instructs the co-processor 20 to transfer data, and transfers data from the data storage unit 2 to the data storage unit 5 on the co-processor side (step A51).

  The processing instruction unit 4 instructs the co-processor 20 to perform data processing. The arithmetic execution unit 6 of the co-processor 20 calculates the sum for the chunks that have been transferred (step A52).

  The process instruction unit 4 repeats steps A51 and A52 until the process instruction is completed for all chunks (step A53: NO). When all chunk processing instructions have been completed (step A53: YES), the co-processor 20 waits for the last chunk processing to end (step A54). When it is detected that the calculation of the last chunk has been completed, the processing instruction unit 4 instructs the co-processor 20 to transfer the processing result. The co-processor 20 reads the final result data from the data storage unit 5 and transfers it to the host computer 10 (step A55).

  FIG. 6 shows the data transfer processing and data processing execution time of each chunk in the above example (optimized). Here, it is assumed that the final transfer time is α. FIG. 6 also shows data transfer processing and data processing execution time when processing is performed in the original order of chunks C0, C1, C2, and C3 for comparison (in-order). As shown in FIG. 6, by performing the processing in the order determined by the order determination unit 3, it is possible to shorten the waiting time for transfer and the processing waiting time rather than performing the processing in order, thereby speeding up the overall processing. it can.

  As described above, according to the present embodiment, when processing data is transferred to a co-processor for processing, the processing data is divided into a plurality of chunks, and each divided chunk is based on the transfer time and processing time. By determining the processing order, it is possible to improve the efficiency of data transfer and the utilization rate of the coprocessor, and to shorten the overall processing time.

  In the above-described processing, the co-processor 20 passes the result to the host side after all chunks have been processed. However, the present invention is not limited to this, and the processing result is passed to the host side for each chunk. Also good. Processing in this case will be described with reference to FIG.

  The processing instruction unit 4 instructs the co-processor 20 to transfer the chunk, reads the chunk from the data storage unit 2 and transfers it to the co-processor 20 according to the order determined by the order determination unit 3 (step A71). The co-processor 20 stores the transferred chunk in the data storage unit 5.

  The processing instruction unit 4 reads the processing content stored in the data storage unit 2, and instructs the co-processor 20 to process the chunk (step A72). The arithmetic execution unit 6 of the co-processor 20 performs data processing instructed by the host for the chunk that has been transferred.

  The processing instruction unit 4 instructs the co-processor 20 to transfer the processing result of the chunk (Step A73). The co-processor 20 transfers the processing result of the chunk to the host, and the processing instruction unit 4 stores the transferred processing result in the data storage unit 2.

  The process instruction unit 4 determines whether or not the process instruction has been completed for all chunks (step A74). If it has not been completed (step A74: NO), the process instruction unit 4 returns to step A71 to perform the process for the next chunk. When the processing instructions for all the chunks have been completed (step A73: YES), the process waits for the end of the last chunk transfer process by the co-processor 20 operation execution unit 6 (step A754).

  When the co-processor 20 finishes transferring the last chunk in the co-processor 20, the process instructing unit 4 performs a process of collecting the result processes for each chunk stored in the data storage unit 2 (step A76).

  When processing as shown in FIG. 7 is performed, the chunk processing order is determined in addition to the chunk transfer time and chunk processing time, the size of the result that can be held on the co-processor side, and the transfer of the result. This is an optimization problem with time as a parameter. In this case, the order determination unit 3 creates an approximate optimal schedule by, for example, a heuristic solution that obtains an optimal schedule by brute force.

  In the above embodiment, in the data registration process (FIG. 3) by the data registration unit 1 of the host computer 10, each function for obtaining the data transfer time and data processing time of each chunk is input. As a modification of the above embodiment, the time acquisition unit 13 of the data registration unit 1 actually acquires and acquires at least one of the data transfer time and the data processing time of each chunk, and the order determination unit 3 based on the actual measurement value. May determine the processing order of each chunk. Data registration processing by the data registration unit 1 in this case will be described with reference to FIG.

  The data registration unit 1 receives input of processing target data and processing contents (step A81). The data dividing unit 11 divides the input processing target data into a plurality of chunks (step A82). The data compression unit 12 compresses each chunk (step A83). The time acquisition unit 13 measures the data transfer time and the data processing time for each compressed chunk (step A84). Specifically, for each chunk, the time from when the data transfer is instructed to the co-processor 20 until the transfer is completed is measured. For each chunk, the time from when the processing is instructed until the processing is completed is measured. The data registration unit 1 stores the compressed chunks, the input processing contents, and the measured transfer time and processing time of each chunk in the data storage unit 2 (step A85).

  Here, both the data transfer time and the data processing time are measured, but only one of them may be measured, and the other may be obtained by a function or the like.

  The data registration unit 1, the order determination unit 3, the processing instruction unit 4, and the calculation execution unit 6 according to the embodiment of the present invention described above read and execute an operation program or the like stored in a storage unit by a processing device such as a CPU. It may be realized by the above, or may be configured by hardware. Only some functions of the above-described embodiments can be realized by a computer program.

  Although the present invention has been described with reference to the preferred embodiments, the present invention is not necessarily limited to the above-described embodiments, and various modifications can be made within the scope of the technical idea.

  A part or all of the above-described embodiment can be described as in the following supplementary notes, but is not limited thereto.

(Appendix 1)
A data processing scheduling device for transferring data to a co-processor for processing;
A dividing means for dividing the data to be processed into a plurality of chunks;
Time acquisition means for acquiring the data transfer time and data processing time of each chunk;
Order determining means for determining the order in which each of the chunks is processed by the co-processor based on the magnitude relationship between the data transfer time and the data processing time of each chunk;
A data processing scheduling apparatus comprising:

(Appendix 2)
The order determining means classifies each chunk based on the magnitude relationship between the data transfer time and the data processing time, and processes the group of chunks having the smaller data transfer time in the order of the smaller data transfer time. Determine the order of each chunk so that the larger chunk group is processed in order of the data processing time,
The data processing scheduling apparatus according to supplementary note 1, wherein:

(Appendix 3)
The order determination means sorts groups of chunks having a smaller data transfer time in ascending order of data transfer time, sorts groups of chunks having a larger data transfer time in order of increasing data processing time, Combine the sorting results of the two groups so that the smaller chunk group comes first, and generate a schedule list of the chunk processing order.
The data processing scheduling apparatus according to claim 2, wherein:

(Appendix 4)
The time acquisition means measures and acquires the data transfer time and data processing time of each chunk,
The said order determination means determines the order which makes a said co-processor process each chunk based on the measured data transfer time and data processing time of each chunk. The any one of Claim 1 to 3 characterized by the above-mentioned. A data processing scheduling apparatus according to claim 1.

(Appendix 5)
There is a data processing scheduling method for transferring data from a host to a co-processor and processing it.
Divide the data to be processed into multiple chunks,
Get the data transfer time and data processing time for each chunk,
Based on the magnitude relationship between the data transfer time and the data processing time of each chunk, the order in which each of the chunks is processed by the co-processor is determined.
A data processing scheduling method characterized by the above.

(Appendix 6)
When determining the order in which each of the chunks is processed by the co-processor, the chunks are separated based on the magnitude relationship between the data transfer time and the data processing time, and a group of chunks having a smaller data transfer time is determined as data. Determine the order of each chunk so that the group of chunks with the larger data transfer time after processing in ascending order of the transfer time is processed in the order of the larger data processing time.
The data processing scheduling method according to claim 5, wherein:

(Appendix 7)
When determining the order in which the co-processor processes each chunk, the group of chunks with the smaller data transfer time is sorted in ascending order of the data transfer time, and the group of chunks with the longer data transfer time is processed with data. Sorting in descending order of time, combining the sorting results of the two groups so that the group of chunks with the smaller data transfer time comes first, and generating a schedule list of chunk processing order,
The data processing scheduling method according to claim 6.

(Appendix 8)
Measure and acquire the data transfer time and data processing time of each chunk,
Determining the order in which each chunk is processed by the co-processor based on the measured data transfer time and data processing time of each chunk when determining the order in which each co-processor is processed by the co-processor. The data processing scheduling method according to any one of claims 5 to 7, wherein:

(Appendix 9)
On the computer,
Split processing to divide the data to be processed into multiple chunks,
Time acquisition processing to acquire the data transfer time and data processing time of each chunk,
An order determination process for determining the order in which each of the chunks is processed by the co-processor based on the magnitude relationship between the data transfer time and the data processing time of each chunk;
A program characterized by having executed.

(Appendix 10)
In the order determination process, the chunks are classified based on the magnitude relationship between the data transfer time and the data processing time, and the group of chunks having the smaller data transfer time is processed in the order of the data transfer time, and then the data transfer time Determine the order of each chunk so that the larger chunk group is processed in order of the data processing time,
The program according to claim 9.

(Appendix 11)
The order determination process sorts groups of chunks with a smaller data transfer time in ascending order of data transfer time, sorts groups of chunks with a larger data transfer time in order of increasing data processing time, Combine the sorting results of the two groups so that the smaller chunk group comes first, and generate a schedule list of the chunk processing order.
The program according to claim 10.

(Appendix 12)
The time acquisition process measures and acquires the data transfer time and data processing time of each chunk,
The order determination process determines an order in which the co-processor processes each chunk based on the measured data transfer time and data processing time of each chunk. The program according to item 1.

DESCRIPTION OF SYMBOLS 1 Data registration part 2 Data storage part 3 Order determination part 4 Processing instruction | indication part 5 Data storage part 6 Operation execution part 7 Data bus 10 Host computer 11 Data division part 12 Data compression part 13 Time acquisition part 20 Co-processor

Claims (6)

A data processing scheduling device for transferring data to a co-processor for processing;
A dividing means for dividing the data to be processed into a plurality of chunks;
Time acquisition means for acquiring the data transfer time and data processing time of each chunk;
Order determining means for determining the order in which each of the chunks is processed by the co-processor based on the magnitude relationship between the data transfer time and the data processing time of each chunk;
A data processing scheduling apparatus comprising: The order determining means classifies each chunk based on the magnitude relationship between the data transfer time and the data processing time, and processes the group of chunks having the smaller data transfer time in the order of the smaller data transfer time. Determine the order of each chunk so that the larger chunk group is processed in order of the data processing time,
The data processing scheduling apparatus according to claim 1, wherein: The order determination means sorts groups of chunks having a smaller data transfer time in ascending order of data transfer time, sorts groups of chunks having a larger data transfer time in order of increasing data processing time, Combine the sorting results of the two groups so that the smaller chunk group comes first, and generate a schedule list of the chunk processing order.
The data processing scheduling apparatus according to claim 2, wherein: The time acquisition means measures and acquires the data transfer time and data processing time of each chunk,
The said order determination means determines the order which makes a said co-processor process each chunk based on the measured data transfer time and data processing time of each chunk. The any one of Claim 1 to 3 characterized by the above-mentioned. A data processing scheduling apparatus according to claim 1. There is a data processing scheduling method for transferring data from a host to a co-processor and processing it.
Divide the data to be processed into multiple chunks,
Get the data transfer time and data processing time for each chunk,
Based on the magnitude relationship between the data transfer time and the data processing time of each chunk, the order in which each of the chunks is processed by the co-processor is determined.
A data processing scheduling method characterized by the above. On the computer,
Split processing to divide the data to be processed into multiple chunks,
Time acquisition processing to acquire the data transfer time and data processing time of each chunk,
An order determination process for determining the order in which each of the chunks is processed by the co-processor based on the magnitude relationship between the data transfer time and the data processing time of each chunk;
A program characterized by having executed.

Images