CN111913881B

CN111913881B - Method for generating I/O trace of application program

Info

Publication number: CN111913881B
Application number: CN202010713191.1A
Authority: CN
Inventors: 谢雨来; 冯丹; 杨震
Original assignee: Huazhong University of Science and Technology
Current assignee: Huazhong University of Science and Technology
Priority date: 2020-07-22
Filing date: 2020-07-22
Publication date: 2022-02-15
Anticipated expiration: 2040-07-22
Also published as: CN111913881A

Abstract

The invention belongs to the field of computer storage, and particularly discloses a method for generating an application program I/O trace, which comprises the following steps: acquiring partial real I/O traces of a target application program, converting each real I/O trace into an input data format for generating a countermeasure network, and performing noise cleaning to obtain an I/O trace data set for training; training a generation countermeasure network based on the I/O trace data set for training to generate a synthetic I/O trace of the target application program; and respectively replaying each I/O trace and each synthesized I/O trace in the I/O trace data set for training, evaluating the accuracy of each synthesized I/O trace according to the replay performance of the I/O trace and the synthesized I/O trace, and screening to obtain the final I/O trace of the target application program. The invention defines a generation countermeasure network architecture to accurately generate the I/O trace so as to collect large-scale I/O trace without influencing the running of a specific application program, and solves the problems that the system overhead of a collection technology depending on a source code is large and the running efficiency of an original application program is influenced.

Description

Method for generating I/O trace of application program

Technical Field

The invention belongs to the technical field of computer storage, and particularly relates to a method for generating an application program I/O trace.

Background

At present, with the rapid development of high-performance computing technology and the development and daily progression of cloud storage and cloud computing technology, supercomputers also play more and more important roles. However, limited by the development of devices, the complexity of optimization algorithms, and the complex interaction behavior of multiple software components, resulting in relatively slow development of I/O performance. Therefore, I/O performance has become a performance bottleneck for many high performance computing systems and parallel big data analytics systems. Most parallel systems find the root cause of system I/O performance inefficiency by analyzing the I/O trace. Therefore, a technique for efficiently acquiring a large-scale I/O trace is imperative.

The existing technologies for acquiring the I/O trace mainly fall into two categories: two types are source code dependent acquisition and source code independent acquisition. Techniques that rely on source code to obtain I/O trace need to run with the original application, gathering detailed I/O access information. The system overhead of the technology is very large, and the running efficiency of the original application program can be influenced, especially for a large-scale high-performance computing system. The technology for obtaining I/O trace independent of source code mainly comprises ScalalIOExtrap, which is used for calculating large-scale I/O trace by using small-scale I/O trace through mathematical derivation and is used for generating trace of systems with different scales under the same workload. The trace generated by the technology is limited to the same workload, and the applicability and the fault tolerance rate are low by using a mathematical derivation mode.

Disclosure of Invention

The invention provides a method for generating an I/O trace of an application program, which is used for solving the technical problem that the collection of the I/O trace in the traditional I/O trace acquisition method has high cost and influences the operation efficiency of the original application program.

The technical scheme for solving the technical problems is as follows: a method for generating an application program I/O trace comprises the following steps:

acquiring partial real I/O traces of a target application program, converting each real I/O trace into an input data format for generating a countermeasure network, and performing noise cleaning to obtain an I/O trace data set for training;

training a generation countermeasure network based on the training I/O trace dataset to generate a synthetic I/O trace for the target application;

and respectively replaying each I/O trace and each synthesized I/O trace in the I/O trace dataset for training, and evaluating the accuracy of each synthesized I/O trace according to the replay performance of the I/O trace and the synthesized I/O trace so as to obtain the final I/O trace of the target application program through screening.

The invention has the beneficial effects that: the method defines a generation countermeasure network architecture to accurately generate the I/O trace, so that the large-scale I/O trace can be collected under the condition that the running of a specific application program is not influenced. Specifically, the countermeasure network GAN architecture is generated using a small amount of real I/O trace training collected in advance, its intrinsic features are learned, and then large-scale trace generation is performed. The method solves the problems that the system overhead of the collection technology depending on the source code is large and the operation efficiency of the original application program is influenced. The characteristic data in the I/O trace needs to be formatted and converted into a type which can be received by the generation countermeasure network. In addition, before training, unimportant characteristic values and abnormal values in the I/O trace need to be eliminated, so that the training efficiency of the generation countermeasure network is improved, and the accuracy of the synthesis I/O trace is guaranteed.

On the basis of the technical scheme, the invention can be further improved as follows.

Further, the I/O trace is used as time sequence data, and the generator for generating the countermeasure network is LSTM.

Further, the discriminator of the generation countermeasure network is LSTM or CNN.

Further, the loss function employed in training to generate the countermeasure network is a cross-entropy loss function or a loss function based on the Wasserstein distance.

The invention has the further beneficial effects that: efficient selection of generators, discriminators and loss functions can ensure accurate generation of I/O traces. And the optimal GAN architecture is selected to generate the I/O trace, so that the characteristics in the trace can be accurately simulated, and the large-scale generation is performed. Compared with a collection technology based on mathematical derivation, the method has the advantages of stronger expansibility, higher fault tolerance rate and wider applicability.

Further, each of the real I/O traces includes five features, respectively, an application specific unit, a logical block address, a transfer size, an opcode, and a timestamp.

The invention has the further beneficial effects that: the core characteristic data in the I/O trace needs to be extracted and subjected to format conversion so as to be converted into a type which can be received by the generated countermeasure network, the training efficiency of the generated countermeasure network is improved, and the accuracy of synthesizing the I/O trace is ensured.

Further, the data format conversion method for the transmission quantity size characteristic is as follows: the transfer size data is divided by 512.

The invention has the further beneficial effects that: since the value range of the transmission size field is very large and is a multiple of 512, the value range of the field is effectively reduced by dividing the field by 512 in order to facilitate the training of the network.

Further, the data format conversion method for the operation code features is as follows: the opcode field is binarized and replaced with either a 1 or a-1.

Further, the noise cleaning includes:

clearing type data except read and write type data in the operation code characteristics; and clearing outliers in the transmission quantity size characteristic and the logic block address characteristic respectively.

Further, after data format conversion is carried out on each real I/O trace, normalization processing is carried out on data corresponding to each feature in all the real I/O traces.

The present invention also provides a computer-readable storage medium including a stored computer program, wherein when the computer program is executed by a processor, the apparatus on which the storage medium is located is controlled to execute the method for generating the application I/O trace.

Drawings

Fig. 1 is a flowchart of a method for generating an application I/O trace according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of a design flow for generating a countermeasure network architecture according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of a training process for generating a confrontation network according to an embodiment of the present invention;

FIG. 4 is a schematic diagram illustrating an I/O trace replay flow according to an embodiment of the present invention;

fig. 5 is a schematic overall flow chart of data format conversion according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

Example one

A method for generating an I/O trace of an application, as shown in fig. 1, includes:

training a generation countermeasure network based on the I/O trace data set for training to generate a synthetic I/O trace of the target application program;

and respectively replaying each I/O trace and each synthesized I/O trace in the I/O trace dataset for training, and evaluating the accuracy of each synthesized I/O trace according to the replay performance of the I/O trace dataset and the synthesized I/O trace so as to screen and obtain the final I/O trace of the target application program.

A generation countermeasure network architecture is newly defined to accurately generate the I/O trace, so that the large-scale I/O trace can be collected under the condition that the running of a specific application program is not influenced. Specifically, the countermeasure network GAN architecture is generated using a small amount of I/O trace training collected in advance, its intrinsic features are learned, and then large-scale trace generation is performed. The method solves the problems that the system overhead of the collection technology depending on the source code is large and the operation efficiency of the original application program is influenced.

Preferably, when the I/O trace is used as the time series data, the generator for generating the countermeasure network is LSTM.

The I/O trace may be treated as time series data. LSTM was chosen as the generator in generating the countermeasure network in view of its outstanding ability to generate time series data.

Preferably, the discriminator that generates the countermeasure network is the LSTM.

The main role of the discriminator in the generation countermeasure network architecture is to discriminate the input value as either the generation trace or the true trace, i.e., to classify both. Considering the ability of CNN to classify and LSTM to process time series data, CNN and LSTM were chosen as discriminators for generating the antagonistic network GAN, respectively. Further experiments tested the effect of CNN and LSTM as discriminators, with LSTM ultimately preferred as a discriminator in the GAN module.

Preferably, the loss function used in training to generate the countermeasure network is a cross-entropy loss function or a loss function based on the Wasserstein distance.

The main role of generating a penalty function in the antagonistic network architecture is to maximize the discriminative power of the discriminator on the generated trace and the real trace. And respectively selecting a cross entropy loss function and a Wasserstein distance-based loss function as the loss functions in the GAN module.

As shown in fig. 2, the specific design flow for generating the countermeasure network architecture may be as follows:

(1) by analyzing the characteristics of the I/O trace, the I/O trace can be found to be similar to the time series data and is the time-stamped sequence data.

(2) The generator is used for generating I/O trace according to input random noise simulation, and the LSTM and the fully-connected neural network are designed as the generator in consideration of the outstanding capability of the LSTM in the aspect of sequence data prediction.

(3) The discriminator is used for discriminating the input value into the generation trace or the real trace, namely classifying the two. LSTM, CNN and fully-connected neural networks were designed as discriminators, taking into account the ability of CNN in classification and the ability of LSTM in time-series data processing.

(4) The main role of the loss function is to maximize the discriminative power of the discriminator on the generated trace and the true trace. A cross entropy loss function and a Wasserstein distance-based loss function are selected.

(5) The generator, discriminator and loss function are combined into a plurality of generation countermeasure network architectures. A total of five groups of architectures are combined, as shown in table 1:

table 1 generating a countermeasure network architecture

Name (R)	Generator	Discriminator	Loss function
				LLGAN	LSTM	LSTM	Cross entropy function
WLLGAN	LSTM	LSTM	Wasserstein distance
				LCGAN	LSTM	CNN	Cross entropy function
WLCGAN	LSTM	CNN	Wasserstein distance
				OGAN	Fully connected neural network	Fully connected neural network	Cross entropy function

(6) By comparing the generation effects of the GAN framework, the optimal generation countermeasure network framework LLGAN is finally selected, namely the generator and the discriminator are LSTM, and the loss function is a cross entropy function.

Efficient selection of generators, discriminators and loss functions can ensure accurate generation of I/O traces. And the optimal GAN architecture is selected to generate the I/O trace, so that the characteristics in the trace can be accurately simulated, and the large-scale generation is performed. Compared with a collection technology based on mathematical derivation, the method has the advantages of stronger expansibility, higher fault tolerance rate and wider applicability.

As shown in fig. 3, a specific process of generating the confrontation network training may include the following steps:

(1) and setting hyper-parameters in the neural network, such as the step size of Epoch and LSTM, the learning rate and the like. Random noise conforming to a normal distribution is generated.

(2) And initializing parameters such as weight, bias and the like in the neural network, and starting training.

(3) Judging whether the Epoch meets the requirement, if so, jumping to the step (8); if not, jumping to step (4).

(4) Random noise is input into the generator, and the true trace distribution is simulated to generate trace through the conversion of the neural network.

(5) And inputting the generated trace and the real trace into the discriminator, and outputting two probability values representing the probability that the input value is the real trace or the generated trace.

(6) And calculating the loss function value of the discriminator and the generator according to the output value of the discriminator.

(7) Parameters in the generator and discriminator are updated based on back propagation of the loss function values. And (4) jumping to the step (3).

(8) And storing the trained network model, and generating the I/O trace directly through the model.

Preferably, before training to generate the countermeasure network, each of the partial real I/O traces is converted into an input data format for generating the countermeasure network and subjected to data cleansing, wherein each real I/O trace includes five features, namely an application specific unit, a logical block address, a transmission size, an operation code and a time stamp.

The core feature data in the I/O trace needs to be extracted and formatted to be converted into a type that can be received by the generative countermeasure network, and then input into the generative countermeasure network to train and generate the I/O trace, where the concepts of features and feature lists need to be explained as follows: for example, if a data set is composed of a plurality of I/O traces, each I/O trace includes a plurality of feature data, a plurality of feature columns exist in the data set, and each element in each feature column is data corresponding to the feature in each I/O trace.

The current I/O trace dataset generally includes many feature rows, but the core feature row is relatively fixed, mainly including the following five features, and the core feature is described below by taking SPC trace as an example.

(1) Application Specific Unit (ASU). The ASU is a monotonically increasing integer starting from zero and used to describe the number of application specific units. The first record in the Trace file does not require the ASU to be equal to zero, but there must be a zero in the Trace file.

(2) Logical Block Address (LBA). The LBA field is a positive integer to describe the ASU block offset for the data transfer of this record, where the size of the ASU block is contained in the Trace file description. The value of this offset may be between 0 and n-1, where n is the capacity of the ASU block.

(3) Size of transmission amount (Size). The Size field is a positive integer that may be zero to describe the number of bytes transferred for this record. The result depends on the I/O subsystem and is a multiple of 512.

(4) An operation code (Opcode). The Opcode field is a separate, case-insensitive character that defines the direction of I/O transmission. It has two possible values: "r", "w", "r" denote read operations, meaning the transfer of data from the ASU to the host. "w" represents a write operation, meaning that data is transferred from the host to the ASU.

(5) A Timestamp (Timestamp). The Timestamp field is a positive real number that indicates the time (in seconds) of each record. The format of this field is "s.d", where "s" denotes the integer part and "d" denotes the fractional part. Both the integer and fractional parts of the field must be present.

It should be noted that, since the I/O trace format is converted to adapt to the generation countermeasure network, when the replay is performed after the generation, the code of the data reading part of the existing I/O trace replay tool needs to be modified, and the accuracy of the generated trace is evaluated by comparing the data volume size and the replay time of the real trace and the generated trace. As shown in fig. 4, the overall flow steps of I/O trace replay are as follows:

(1) and reading the real trace or generating the trace and storing the trace into an array. Carrying out the step (2) and the step (5) in parallel;

(2) changing a source code of a data reading part in the Btrrecord tool to read the used I/O trace data set;

(3) generating a replay file corresponding to the specified trace data set by using a Btrrecord tool;

(4) and playing back the generated replay file by using the Btreplay tool. Jumping to the step (7);

(5) changing the source code of the data structure part under the run mode of the Docker-replayer tool to replay the used I/O trace data set;

(6) and starting multithreading for playback by using a Docker-replayer tool. Jumping to the step (7);

(7) and integrating the replay results of the two tools, analyzing the performance of the real trace and the generated trace, and evaluating the accuracy of the generated trace.

Preferably, the noise cleaning comprises: clearing type data except read and write type data in the operation code characteristics; outliers in the transport size and LBA characteristics are cleared, respectively.

For the accuracy of the synthesized data, the unimportant characteristic values and abnormal values in the I/O trace need to be cleared before training, for example, for the operation code field (operation code characteristic) in the I/O trace, the operation code field is mainly concerned to be data of read and write types, so for a part of I/O trace, the operation code field can be cleared for other types of data; for the traffic size field (traffic size characteristic), there may be some discrete points (traffic maximum or minimum values) for which accurate identification and clearing is required.

For an I/O trace data set, a feature column is generated by adopting a generation countermeasure network, and the format of the feature column needs to be converted so as to be in accordance with the vector type which can be identified by the generation countermeasure network, wherein different processing modes are adopted for different feature columns, the data format characteristics of each feature column are considered, and two features of operation codes and transmission quantity are mainly processed.

Preferably, the data format conversion for the characteristics of the transmission quantity size is implemented as follows: the transfer size is divided by 512.

Since the value range of the transmission size field is very large and is a multiple of 512, the value range of the field is effectively reduced by dividing the field by 512 in order to facilitate the training of the network.

Preferably, the implementation manner when the data format conversion is performed on the operation code features is as follows: the opcode field is binarized and replaced with either a 1 or a-1.

Since the operation code field has only two values, the operation code field is binarized to convert "r" to-1 and "w" to 1.

As shown in fig. 5, the overall flow of data format conversion may be as follows:

(1) an array storing I/O traces is initialized. The size of the array is n x, wherein n is the number of rows of the trace file, and x is the number of characteristic columns in the trace array;

(2) reading the I/O trace file line by line, wherein each record is equivalent to a line value of an array;

(3) judging whether the tail of the file is reached, if so, performing the step (6); if not, performing the step (4);

(4) for example, for the transmission size, it is determined whether it is a multiple of 512 or 1024, and if so, the range is reduced by dividing by the corresponding size. For example, for the operation code field, it is binarized according to "r" or "w" and converted into-1 or 1;

(5) and (4) storing the preprocessed vectors into an array, and performing the step (2).

(6) After the trace file is completely read, the array columns for storing the I/O trace are normalized, and the numerical values are scaled to the range of [ -1,1 ].

Preferably, after the data format conversion is performed on each feature column, the data of each feature column is normalized.

And dividing the converted characteristic column vectors into columns for normalization, and scaling the numerical value of each column into the range of [ -1,1], so that the training and the generation of the network are facilitated.

In general, the present embodiments provide a method for generating I/O trace based on generation of a countermeasure network, i.e., a record of I/O requests accepted by some systems running for multiple days of disk. The method includes data preprocessing, noise cleaning, GAN training and I/O trace replay. In particular, a GAN architecture is defined to learn the characteristics of an I/O trace dataset, and then large-scale accurate generation is performed. The data preprocessing specifically includes receiving an original trace data set and converting the original trace data set into a data format which can be received by the GAN. The noise removal specifically removes characteristic values and abnormal values which are not required to be extracted from the data. The GAN mainly comprises a generator and a discriminator, wherein the generator inputs a noise value and outputs a generated I/O trace; the discriminator inputs the real trace and generates the trace, and outputs a probability value of the input value being the real trace or generating the trace. Loss function values of the generator and the discriminator are calculated according to the output value of the discriminator, back propagation is carried out, parameters are updated, and the accuracy of generating the I/O trace is improved. The I/O trace replay is to replay the generated trace and the real trace, and the accuracy of the generated trace is verified by comparing the performances of the generated trace and the real trace.

Example two

A computer-readable storage medium comprising a stored computer program, wherein when the computer program is executed by a processor, the computer program controls a device on which the storage medium is located to execute a method for generating an application I/O trace according to the first embodiment. The related technical solution is the same as the first embodiment, and is not described herein again.

It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims

1. A method for generating an application I/O trace, comprising:

respectively replaying each I/O trace and each synthesized I/O trace in the I/O trace dataset for training, and evaluating the accuracy of each synthesized I/O trace according to the replay performance of the I/O trace and the synthesized I/O trace so as to obtain the final I/O trace of the target application program through screening;

each real I/O trace includes five features, which are an application specific unit, a logical block address, a transfer size, an opcode, and a timestamp, respectively;

the noise cleaning includes: clearing type data except read and write type data in the operation code characteristics; and clearing outliers in the transmission quantity size characteristic and the logic block address characteristic respectively.

2. The method of claim 1, wherein the I/O trace is used as time series data, and the generator for generating the countermeasure network is LSTM.

3. The method of claim 2, wherein the identifier for generating the countermeasure network is LSTM or CNN.

4. The method of claim 1, wherein the loss function used in training the generation of the countermeasure network is a cross-entropy loss function or a Wasserstein distance-based loss function.

5. The method according to claim 1, wherein the data format conversion of the transmission size characteristics is performed by: the transfer size data is divided by 512.

6. The method according to claim 1, wherein the opcode features are converted into data formats in a manner that: the opcode field is binarized and replaced with either a 1 or a-1.

7. The method according to claim 1, wherein after performing data format conversion on each of the real I/O traces, performing normalization processing between data corresponding to each feature in all real I/O traces.

8. A computer-readable storage medium, comprising a stored computer program, wherein when the computer program is executed by a processor, the computer program controls a device on which the storage medium is located to execute a method for generating an application I/O trace according to any one of claims 1 to 7.