Why do Alibaba Cloud and Amazon "volume" cloud computing infrastructure?

I do n’t know if the concept of the Yuan universe is a pseudo -pseudo -concept, but today the world has become a digital world, but people ’s lives, work, and study are inseparable from various digital technology blessings, and the bottom of the digital world is based on the basis of the calculation foundation of the cloud Facilities, more specific, is the computing infrastructure of computing resources such as chips, storage, and network computing resources.

Calculation is the basis of the operation of the digital world, just like the sun to the earth.

Before the emergence of cloud computing, the underlying computing center of the computer is CPU and CPU, storage, network, etc. around the CPU center, and other computing resources allocated by CPUs. After the emergence of cloud computing, the traditional computing architecture system became stretched. Because of its facing physical equipment, cloud computing has gone through two stages of distributed+virtualization and resource pooling. In fact, it has been built on the bottom layer. A super computer can be rented on demand for resources after paying. However, this model is not considered by the designer of the former computing architecture system. There are many problems. After the advent of digitalization, more, larger, and more broken data, real -time, massive, and variable computing scenarios, new requirements for calculations, especially high -performance, dense, high toughness, low -carbonization, traditional CPUs The calculation architecture is as difficult as the "Little Mara Big Car".

Over the years, many cloud computing giants have been exploring new computing systems applicable to or exclusive cloud computing systems, and many cloud computing giants from abroad, and domestic cloud computing giants come from Alibaba Cloud.

Alibaba Cloud CIPU enters the game, CPU is no longer the "center" of calculation

On June 13, Alibaba Cloud released CIPU to the outside world, which is called the Alibaba Cloud infrastructure processor (Cloud Infrastic Processing Units), which is the computing control and acceleration center dedicated to the new cloud computing center.

In the traditional computing architecture, the CPU carries this function, which is responsible for core computing and controlling the control and control of resources such as networks and storage. The CPU is the master, other resources are from; CPU is the center, and other resources are peripherals.

Relying on the CPU's computing architecture, it is difficult to support the current cloud computing needs. On the one hand, more and more data -intensive computing of cloud computing response is increasing. The architecture centered on CPU leads to a large delay between computing and network transmission. On the other hand, the amount of data migration in the data center increases, and the architecture centered on the CPU cannot provide high bandwidth. The CPU restricts the low latency and high bandwidth capacity of cloud computing, which has also caused many common applications to achieve greater difficulty in real -time audio and video communication, Yuan cosmic XR, and rising autonomous driving. Developers need to realize ways to realize these applications. The development cycle, development difficulty, and computing costs have increased significantly. The industry also has a PAAS service provider specifically solving the gap between cloud computing platforms and application scenarios.

The bell must also be tied to the bell. To solve the problem of CPU -centric architecture, it is necessary to reorganize and reconstruct this system. Alibaba Cloud's latest solution is to start another stove, move the control center from the CPU inside the server to the CIPU outside the server.

In the CIPU, the function of traditional CPUs is only its functional set. It supports traditional CPUs, which is inserted and played, virtualized, and hardware reinforcement. The integrated RDMA network acceleration is integrated on the network resources, and all support the functions of virtualization, forwarding acceleration, and hardware encryption.

Under the CIPU architecture, the downward access is physical computing, storage, network resources, rapid cloudization and hardware acceleration; connecting the flying sky cloud operating system, through the large -scale application of RDMA network technology, allowing the access cloud to access the local area to access the local area The hard disk is fast. When the data center or cloud computing center is applied to CIPU, it can solve the core issues such as bandwidth, delay, performance, and energy consumption that are currently facing, and then better support different cloud business, help the industry digitize the upgrade Support the emergence of scientific and technological innovation applications.

Look at a set of official data.

A new generation of cloud computing architecture system based on CIPU and Feitian shows superior performance in the calculation test of the core scenarios such as general computing, big data, and artificial intelligence:

In the field of general distributed computing, Redis performance has been improved by 68%, mysql has increased by 60%, Nginx increases by 30%;

After the high -throughput Internet business, after clouds, the cluster throughput of the self -built physical machine was increased by 30%, and the delay of the business peak period decreased by 90%;

In the dual -dense scenario of big data and AI computing and data, compared with the traditional TCP network, the elastic RDMA high -performance network throughput has increased by more than 30%;

In terms of Yunyuan, the container startup speed is 350%faster, and 3,000 elastic container instances can be pulled up in the Serverless scenario in 6 seconds.

Cloud computing manufacturers have claimed that their computing structure has a significant improvement, and the final inspector is developers. The new generation cloud computing architecture system of Alibaba Cloud CIPU+Feitian is the final effect, and the market will give answers because if this architecture can really overcome the lack of CPU -centric computing architecture on cloud computing, it can achieve such a powerful effectiveness improvement For developers, it will mean that better, practical, and affordable resources will have a stronger cost -effective for Alibaba Cloud. This is still the core competitiveness of the cloud computing market. "Feitian+CIPU", do you want to do the "iOS+A" partner in the new era of cloud computing?

As we all know, the calculations in any scenario are inseparable from the effective combination of soft hardware. The performance of the chip must have software system algorithms to eat. Moore's law must be driven by market demand. Some chip performance no longer follows that the law of Moore is not the technical ceiling, but for the market, it has excessive performance. And the chip performance is eaten, there are two ways:

One is ecology. The most classic is the "Wintel camp" built by Window-Intel. At that time, Intel CEO Andy Gruf continued to upgrade the chip performance. Eat, this is called "Andybier's Law". While the Wintel Alliance allows Intel chips to continue to improve, it restricts the development of other chips. The mobile era Qualcomm (and ARM camp) and Android also form a similar league.

The other is self -produced and self -selling, such as Apple A processor, Huawei Hisilicon, because it is its own product, combining the application scenario with the processor can exert the strongest performance. Now the Apple A series, the M series processor Intel and Qualcomm have stressed. Alibaba Cloud's "Feitian+CIPU" is the "iOS+A processor" partner mode in the field of cloud computing. Both iOS and A processors are born to mobile devices. Feitian+CIPUs are born for cloud computing.

In 2009, before Alibaba Cloud launched, Ali decided to independently develop a large -scale distributed computing operating system "Feitian". This is a operating system that is specifically for cloud computing. The world's million -level server controls it. Now that Flying and CIPU software and hardware, Alibaba Cloud has redefined the computing architecture of cloud computing.

CIPU is essentially encapsulated different capabilities -computing, storage and network resources, as well as the corresponding acceleration technology industry for many years.

It is not uncommon to encapsulate. While the CPU is constantly increasing the core number, it is also packaging more capabilities. For example, the AI ​​chip that has been popular in the past two years will integrate the GPU and other computing units to strengthen the localized AI computing. Intel is in it A few days ago, reiterated again in the next few years to integrate high -performance CPUs and GPUs to a chip with updated manufacturing processes and packaging technology.

In the field of cloud computing, the exploration of giants has also had a history of many years.

As Zhang Jianfeng, the president of Alibaba Cloud Intelligent, was summarized by the CIPU, the development of cloud computing has gone through two stages in more than 10 years. The first stage is distributed and virtualized technology to replace large -scale and small machines. , Enterprises no longer need to build self -built machine room maintenance. From buying the rent on demand to expand, the computing resources used are still a console. In the second stage, resource pooling technology occurs. By calculating the storage separation architecture, computing, storage, and network resources are pooling separately, and the bottlenecks of scale and stability can provide large -scale cloud computing services.

The two stages of "distributed+virtualization" and "resource pooling" are optimized to computing, storage, and network resources through software definition methods. Enterprises use software to define the advantages of the aggregation of computing resources to exert scale effects, just like Alibaba Cloud Flying. However, this model has been difficult to adapt to this era, because the needs of calculation have changed, more industries, more customers, more businesses, new scenes (such as cloud border fusion, audio and video live broadcast, Yuan cosmic XR, etc. Wait) all clouds, the result is more massive dense data and corresponding AI computing requirements. These are high in low latency, high bandwidth, and low -carbonization of cloud computing, and it is difficult to meet the traditional architecture.

Alibaba Cloud has long realized that the traditional architecture centered on CPU supports cloud computing will only become increasingly difficult. Therefore, in 2015, a special technical tackling team was established. Based on the CPU+FPGA scheme to realize the support of nude metal virtualization, it has made naked metal servers that have beyond the physical machine. Since then, the Shenlongyun server iterates to the fourth generation. promote.

Cloud computing is still running on the host after virtualization. The host must allocate some CPUs and memory resources to run the DOM0, which is the privilege virtual machine (the manager and controller of other virtual machines), which leads to 10%-30%calculation Resources cannot be sold, and the cost of cloud computing is increased. This part of the cost is "data center tax".

However, there are still many problems that are difficult to solve by Shenlongyun server alone. Customers have higher requirements for high bandwidth, low latency, and low -carbonization. The cost of Moore's law is huge, and increasing the frequency of calculation will increase the fever and power consumption and increase the operating cost. These do not meet the core interests of customers. When Alibaba Cloud's core technologies such as Shenlong server, elastic RDMA, and self-developed RISC-V instruction set chip, the global cloud computing giant is naturally not idle.

After Amazon acquired the Israeli chip company Annapurna Labs in 2015, it developed a customized chip for cloud computing infrastructure. In 2018, the first generation of Amazon Graviton processor was released. Iterations, Graviton 3, which uses a 5nm process in December 2021, has a significant improvement in performance, energy consumption and other performances, which can better support workloads such as scientific computing, machine learning, and media coding. However, the direction of Amazon's efforts is still the CPU itself, and the performance of this traditional computing architecture system is better through customized means. Microsoft was also exposed to develop a customized chip to the cloud computing server. This year, Mike Filippo, an apple semiconductor expert, was dug to engage in the development of processors.

Google's solution is to start another stove. It no longer adopts universal chips such as CPU and GPU. It does not use FPGA technology. Instead, it is customized for a special chip suitable for specific computing scenarios: TPU chip, which serves Google AI computing. The full name of the TPU is Tensor Processing Unit, that is, the tension processing unit, which is tailor -made for Google Machine Learning Platform TensorFlow. Compared with the general chip, it is more suitable for running neural networks. It is reported that the Google TPU chip surpasses the Intel to Qiangqiang CPU and the Geezo -Qiangqiang CPU and Nvidia GPU is a number. In addition, Google has a video decoding chip created by video applications such as YouTube, such as ARGOS. For a new chip customized in a specific computing scene, this is Google's approach. It can better meet the computing needs in some scenarios, but failed to do it once and for all.

Amazon's customized chip model based on ARM architecture failed to solve the "CPU -centric computing architecture in supporting cloud computing scenarios", and Google's scenario -based customized model is difficult to solve the massive and complex general -purpose computing problems in this era.

Amazon and Google's approach has been doing it a few years ago, and also launched AI chip containing light 800 and CPU processor Yitian 710. It is precisely because of the accumulation of Shenlongyun server, elastic RDMA, and flat -headed brother chip that Alibaba Cloud can launch a new architecture of "CIPU+Flying" today. In combination with the software and hardware of Feitian System, the three major resources (computing, storage and network) in depth are integrated to achieve higher performance, lower latency, larger bandwidth, and lower power consumption computing, adapting to high -performance calculations, real -time calculations , Data -intensive calculation and other mainstream scenes.

Although the CIPU carries the role of "computing control and acceleration core" of future cloud computing, this is more like a decentralized architecture. Relying on it means that there will be no bottlenecks. The CIPU architecture can also support different chips such as CPU, GPU, as well as different architectures such as ARM and X86, making different computing resources or system complementarity.

Back To Basic, how does Alibaba Cloud return to the essence of cloud computing?

When releaseing CIPU, Zhang Jianfeng said that Alibaba Cloud's most important strategy in 2022 is "B2B", which is "back to basic", which returns to the essence. Growth to the pursuit of high -quality growth is the "return technical nature" that no longer pursue the first scale.

The release of CIPU shows that Alibaba Cloud's heavy position computing power is "calculated" to "calculate" the essence of cloud computing, that is, the customer -centric providing a more extreme computing power service, and it is necessary to do this. Because technology is the root of cloud computing, resources, channels, services, brands, etc. can only be branches and leaves. Only the roots are deep, can it be leafy. How deep is Alibaba Cloud's Back to Basic.

Zhang Jianfeng said that Alibaba Cloud would "adhere to the long march of technology". This is because when Alibaba Cloud was established, it was a technical business. It died in business and used technology to drive business. From going to IOE to Feitian system to Jianzhong to Jianzhong, the system of systematic layout of the Dharma Institute is the frontier of the foundation of the Dharma Institute, and the technical layout is Alibaba Cloud's snowy mountains and grassland.

In fact, CIPU is not designed from Alibaba Cloud from 0 to 1, but is based on the core technologies such as Shenlong, Elastic RDMA, and chip that have been developed by self -developed self -developed for many years, and continuously deepen the result of vertical integration. Like Apple's iPhone, the previous generations used their own systems, but it was Intel's processor. It was not until 2010 that the iPhone 4 was launched on the A4 processor. This was the beginning of the popularization of the iPhone. The era of mobile Internet really came. The self -developed processor and the "iOS+A" computing architecture. Apple allows the mobile computing to complete the "shadow" of PC computing. In terms of energy consumption and other dimensions, there is no shortage of defects, and once and for all solved the underlying problem facing mobile devices. In the same way, Alibaba Cloud also wants to rely on underlying technology to overcome the architecture that has always faced cloud computing. Technical problems are solved with technology, which is also the route that Alibaba Cloud has always adhered to. When the independent research and development system was decided in 2009, Alibaba Cloud had many open source cloud platforms to choose from. Considering the time, cost, and risk dimension, using a mature open source system is the best choice. The technical route, independent research and development, Jiang Jiangwei, the head of Alibaba Cloud Technology R & D, later resumed the media: "If it is not independent research and development, we cannot deal with the peak of the transaction of double 11325,000/s."

The cloud operating system is a magnificent project, and Ali is full of bumps. After three years of technical research, several times of overwhelming, Feitian and Alibaba Cloud finally ushered in Dacheng. In 2013, Alibaba Cloud released the Flying 5K cluster, becoming the world's first cloud computing manufacturer to mobilize more than 5,000 units. Flying can be connected to a million -class server into a supercomputer all over the world. Based on Feitian, Alibaba Cloud can provide customers with the world's leading computing capabilities, and it is ease of time when extending to data centers and smart platforms. Now combining with the CIPU to form a new cloud computing architecture, it will further enhance the core competitiveness of Alibaba Cloud.

Feitian-style technology self-developed routes work again and again, such as the self-developed database system, such as the RISC-V processor of Pingtou "Sword", the Xuantie 910, and the cloud AI reasoning chip contains 800 light, such as Shenlongyun server. The "foundation" or "basic" of technology is the key to Alibaba Cloud's riding dust. Gartner data shows that in 2021, Alibaba Cloud ranked third in the world in the global cloud computing IaaS market share, which has achieved share growth for six consecutive years; the Asia -Pacific market is first with a market share of 25.53%.

Not only Alibaba Cloud, today 3A (Amazon AWS, Microsoft Azure, and ALIYUN) Cloud computing header players in the bottom of the technique of the bottom of the technology, because everyone knows that technology is the only decisive power in the cloud computing long -distance running. While Alibaba Cloud is leading the core technology, it will also help China to have the right to speak core technology on the cloud computing track, as Zhang Jianfeng said: "Cloud computing is getting closer to the next era -a new definition of architecture, a brand -new one Software interface, hardware acceleration. We missed the PC era, but everyone started the same time in the cloud. Now it is the window period for redefining the cloud. If we define it, China can have its own place in the next technical era. "

- END -