 Today, I wanted to share with you the ATUN, which is an AI-based automatic parameter tuning system. I will introduce ATUN from the following four aspects. First, I wanted to introduce what is the performance tuning. This is a simple example for matrix multiplication of 4,800 by 4,800. If we write it in Python language, it takes more than 660,000 seconds to run. However, if we use C language to write it, it only takes about 600 seconds to run. Then if we optimize by multi-switch parallel computing, then we can only take about 37 seconds. Then if we use other optimization algorithms such as vector instructions, it only needs to take 1.99 seconds. This simple example gave me a, we can say, let our Hoji room to improve our performance. Then I will give an instruction about ATUN. As we can say, operating systems have a large number of parameters and complicated correlations between links. However, those tunable parameters control all aspects of the operating system. For example, in Spark, there are so many tunable values. If we use the default values, they can't get the best performance. Now, we want to tune this parameter. For example, there are several advantages. For example, you can, the execution time of Spark can drop 10 times. And the system resource utilization could up to 10 times. And we can also reduce some errors such as OEM and timeout and so on. However, tunable parameters have several challenges. The first one is that there are many parameters in a single system. For example, in Spark, there are more than 200 parameters. Then the second one is the large number of systems. For example, there are so many applications such as the vertical applications, analytics and visualization systems. Different systems need different kinds of parameters. Then the second one is that there are diverse workloads on the system complexity. For example, even when we use Spark, there are all kinds of workloads. For example, there are including the ML labs, streaming, circle, and graph processing and so on. Different workloads also need a different kind of parameters. However, if we're using manual tuning to tune these parameters, there are several disadvantages. For example, it requires long tuning time and it requires individual appearance and has high level costs. However, if we use AI to automate the tuning, there are several advantages. For example, it only needs short tuning time and there is no need for individual appearance. And it only needs machine costs. Therefore, the object of ATUN is that giving an workload, we want to use AI to find the optimal parameters for giving workload and achieve the best performance. Now, let's see the related works. As we can see, tuning is a constant topic. In the 1980s, large manual tuning had been used. In the 2000s, statistical and hierarchical methods are widely used. Recently, in the 2010s, traditional machine learning methods are also widely used. Recently, deep learning and reinforced learning methods are also widely used. The following table gave the performance of all the load methods under several dimensions, such as quality efficiency and to deal with three challenges. As we can see, there is no single kind of method performed best under all the dimensions. Now, we want to use AI to perform the automatic tuning. Now, several advantages. The first one is that for standard coverage, we can leverage a variety of technologies to optimize several layers. The second advantage is the intelligent operation. We can implement tuning opportunities in efficient and intelligent manner. The third one is the target orientation. We can turn the system based on objects and constraints. The third one is the full automation. We can conduct the entire tuning plus schedule automatically. Now, we design an A8 tune, which we use AI in the optimal performance of the system. First, we need to collect all kinds of data, such as system metrics. For example, CPU utilization, and other data, such as framework specific and application specific metrics, and the target metrics. Then, all the metrics are sent to the A8 tune. A8 tune can make intelligent decision making and learn how to put an optimal load, such as system load, framework specific load, and application specific load. These of the magical logs can make the performance of the system better. Now, we will give an architecture of A8 tune. A8 tune analyze the research usage, such as computing, storage, and network of applications, or clouds, support parameter tuning for mainstream applications in the industry, and improve tuning efficiency. The architecture of A8 tune includes three parts. The first part is intelligent decision making, which includes online static tuning, and offline dynamic tuning. And the second part is system characterization, which includes the data platform, feature engineering, training, and inference. The last one is the system interaction, which includes the system full stack monitoring and configuration service. Now, I will give two parts of the offline, including online static tuning and offline dynamic tuning. Now, I will give a detailed introduction about online static tuning. The scenario of online static tuning is the common uses and applications that are needed to be online at all time. The idea is that A8 tune detects the current application workload, matching it to a lower workload based on the classification model, and then outputs empirical parameters. The online static tuning has several key technologies. The first one is the important feature analysis. A8 tune can automatically select important features to characterize applications accurately, and it can have two layer classification, which can accurately identify the current workload. And the third one is the workload change detection. It can automatically identify the application workload changes and implement adaptive optimization. The second one is offline dynamic tuning. Offline dynamic tuning is for advanced users who have high performance requirements. The idea is that offline dynamic tuning includes three parts, include target device, curia, and workloads. The idea is that curia sets parameters for the target device, obtains feedback performance indicators, and it continuously obtains optimal parameters. Therefore, offline dynamic tuning also includes three key technologies. The first one is the important parameter selection. It can automatically select important parameters to reduce such space and improve training efficiency. The second one is optimization algorithm conjunction. It can allow users to select the optimal algorithm based on application scenarios, parameters, types, and performance requirements. The third one is logic-based conjunction. We can add a counter-organized territory text and an optimal parameter to the logic-based to improve subsequent training efficiency. Now, I will give an ATOIN implementation framework. The framework includes three parts. ATOIN server, ATOIN client, and ATOIN engine. The ATOIN server includes data collection tool and system configuration engineering. The ATOIN server must be installed on a target device. The ATOIN client is used to interact with the user and output the results. The ATOIN engine includes all algorithms of ATOIN, such as classification models, key parameter selection models, model training, and automatic parameter training. The three parts can be installed on different machines and they can communicate through the GRPC. Now, we will give several training results. For example, in Spark, there are several capable parameters. If we use that default value, they can get the best performance. Now, we want to use ATOIN to select the best values for the given workload. For example, after ATOIN training, the performance is improved by 30% over the default set-time configuration and 5% over the configuration optimized by professional engineers. The ATOIN efficiency is 5 times than that of manual training. Now, there are also several other training results. For example, we optimized the ETCD set-time with ATOIN in creating performance by more than 10% the results were displayed in the open-or-wire application protein condition. And we also optimized MySQL. MySQL is optimized in six different scenarios open-or-wire and not operate with mature machines with four cores and 60 get-byte memory as well as 32 cores and 64 get-byte memory. In OLTP in static scenario, the throughput is improved by 5 times. In the future, ATOIN will break performance bottlenecks. There are two kinds of improvements. For example, we want to from offline to online. There are several trends. For example, continuous performance training capabilities for online application to deliver performance as a service. For example, there are also several systems. For example, our ATOIN provides online parameter training for cloud databases. And the graduate focus on resource queue training for online applications. However, there are also switch changes. For example, the first change is called start. How do we deliver optimal performance in absence of history data when a new application goes on? The second change is adaptability. How can the cost of online training be reduced when a model is trained against online applications? That changes dynamically. The third one is security. How do we keep the performance of the immediate queueing results above the preset threshold to ensure high system availability during queueing? The second one is from yield mode to color mode. There are also several trends. For example, utilizing AI technologies to save the form, commission or certain configuration optimization to certain design optimization. There are also several papers. For example, LingOS uses AI to reconstruct the color block layer and predict the IO electricity boosting SDI performance. KML is another paper which builds an AI model in the color space to implement load-aware data perfection policies. However, there are also switch changes. The first challenge is computing overhead. How can model complexity and performance be balanced with the limited computing resources available in the color space? The second challenge is timely list. How do we infer and secure the highly dynamical optimization processes? In my second, the second one is university. How do we construct a universal framework for different optimization types that involve data collection and optimization algorithms for various color modules? Here is an A2N repository. Everyone can get a code of A2N. Please feel free to contact me if you have any questions. This is my email. Thank you.