Ali hardware research and development, there are big moves.
Just now, Alibaba Cloud officially released a new processor: CIPU.
Not only is the architecture fully self-developed, but it also claims to “replace the CPU to become the core hardware of a new generation of cloud computing”!
Cloud computing has been in operation for so many years, and the CPU has still firmly occupied the “C position” in the data center.
Just last year, Ali also made great efforts to launch the 5nm server CPU Yitian 710.
Why is this suddenly breaking with tradition?
CIPU, this new face with one more I than CPU, what is the origin?
What exactly is a CIPU?
The full name of CIPU is Cloud Infrastructure Process Units, which meansCloud infrastructure processor.
As can be seen from the name, this is a cloud processor, which is specially used to connect the hardware in the server and the virtualized resources on the cloud.
△CIPU Architecture Diagram
According to Alibaba Cloud, the reason why the CIPU is used to replace the CPU-centric architecture is to better “squeeze” server hardware, obtain more virtualization resources, and make the existing resources more convenient to use.
In terms of software, the CIPU is connected to the Feitian cloud operating system to more efficiently complete the scheduling and scheduling of virtualized resources;
In terms of hardware, Feitian operating system can quickly manage the physical equipment of the data center through the CIPU, and accelerate the network and storage hardware, so that it will not only waste the computing power of the CPU, but also enhance the network and storage performance.
In terms of function, it has four characteristics:
The best carrier for cloud native, that is, each bare metal system can run 2000 containers, and uses sandbox container technology to provide more secure isolation for containers, and the link startup speed is within 50ms;
The chip directly implements the IO engine, in which the storage I/O operations can be performed 3 million times per second, the network I/O can be up to 50 million packets per second, and the storage long tail latency is reduced by 50%;
Chip-level security reinforcement, that is, the encrypted data can be unloaded at high speed, and the immutability of the chip-level hardware can be mapped to the software;
Enhanced converged network, that is, under the blessing of RDMA technology, the network delay can reach a minimum of 5 microseconds, and the bandwidth can reach a maximum of 200GB.
In terms of performance, it has brought a lot of improvements to computing, storage and network resources.
In terms of computing, the CIPU can quickly access the Dragon Cloud servers of different types of resources, reduce the virtualization consumption of a single container by 50%, and the startup speed is 350% faster. Taking running some databases and servers as an example, the performance of Nginx is improved by 89%, Redis is improved by 68%, and MySQL is improved by 60%. In addition, it is also improved for AI and big data scenarios.
In terms of storage, the CIPU can perform hardware acceleration on the block storage access of the storage-computing separation architecture. The storage latency is as low as 30 microseconds, and the bandwidth is up to 200 Gbps. It supports NVME shared access to cloud disk block storage, Oracle RAC, SAP High-availability databases such as Hana seamlessly migrate to the cloud.
On the network, the CIPU performs hardware acceleration on the high-bandwidth physical network, the basic bandwidth reaches 200GB, and adopts the self-developed RDMA-Solar protocol, which reduces the network latency to 16us, which is 30% higher than the cluster throughput of the self-built physical machine. 90% drop in latency during peak business hours.
Interestingly, Alibaba Cloud’s CIPU processor has actually been polished internally for several years.
Its original “inspiration” came from a device called Shenlongka inside Alibaba Cloud.
Shenlongka was born in 2017. It is similar in function to a platform called Nitro released by AWS (integrated with functions such as virtual machine monitor, out-of-band management, etc.), even a little earlier than AWS released.
After several rounds of iterations, the Dragon Card gradually added more capabilities such as scheduling, hardware acceleration, etc., and finally gave birth to the prototype of the CIPU, and then continued to improve this product.
On the one hand, according to Huang Ruirui, director of Alibaba Cloud Intelligent Cloud Architecture, in the past few years, the CIPU has been subjected to a performance & stress “test” like the “Double Eleven”.
On the other hand, many Alibaba Cloud customers have more or less used CIPU-based cloud computing services. Although customers may not have a direct perception of the underlying hardware layer, Alibaba Cloud’s network, storage and other performance have indeed been on the rise in recent years. For example, not long ago, Alibaba Cloud became the only one in China to be rated as one of the top ten computer network research institutions in the world. Chinese companies.
Looking at it now, the emergence of CIPU has indeed broken the “bottleneck” of cloud computing again and improved the overall performance by a large part.
However, to overthrow the idea of CPU “C position” in cloud data centers, it is not the Alibaba family.
Among products with similar concepts on the market, what is the difference between CIPU and IPU and DPU?
Why CIPU?
To clarify this matter, we still have to start from the development history of cloud computing technology.
The development of cloud computing technology in the past ten years can be roughly divided into two stages.
In the first stage, driven by distributed technology, Internet companies began to migrate their business from mainframes to distributed systems, laying the foundation for a distributed architecture.
In the second stage, resource pooling technology emerged. This technology realizes the unified scheduling and arrangement of resources through the separation of computing and storage, making elastic computing possible.
For users, this also means a great improvement in cloud computing reliability and availability.
In both phases, the computing architecture is CPU-centric.
However, when cloud computing develops to today, there are more and more data-intensive scenarios represented by big data applications, and this CPU-centric architecture begins to expose shortcomings:
First, a CPU-centric architecture results in a large delay between computation and network transmission.
Secondly, the increase of big data applications has led to an increase in the amount of data migration within the data center, and the CPU-centric architecture cannot provide high bandwidth.
Furthermore, taking Alibaba Cloud as an example, it manages millions of servers in 27 countries and regions and 84 availability zones around the world. But CPU-centric architectures struggle to address the complex management issues of such hyperscale infrastructures.
In this way, the solution also points in one direction:Break the traditional cloud computing architecture centered on CPU and define a new generation of cloud computing basic technology.
And this is also the latest technology hotspot that major manufacturers are chasing.
For example, NVIDIA’s DPU (Data Processing Units), once released in October 2020, caused a heated discussion in the industry.
As the name implies, DPU focuses on solving the problem of data migration bandwidth. As an integrated acceleration platform, the DPU can offload critical networking, storage, and security tasks from the CPU, reducing CPU overhead.
Lao Huang said at the time:
The data center has become a new type of computing unit, and DPU is an important part of it. The combination of CPU, GPU and DPU can form a fully programmable single AI computing unit, providing unprecedented security and computing power.
And Intel followed closely behind and proposed the concept of “IPU” (Infrastructure Processing Units).
Compared with the DPU, the IPU emphasizes the ability to virtualize and cloud. Through network virtualization, storage virtualization, network storage management, and security functions, it accelerates the network infrastructure and releases CPU cores to improve application performance.
Although there is a slight difference in concept, both DPU and IPU want to replace the CPU and become the core hardware of the data center through software definition + hardware acceleration.
It can also be seen that the CIPU launched by Alibaba Cloud this time is more like a combination of IPU and DPU. It can not only manage and control the data center through cloud virtualization, but also solve the problem of data migration bandwidth.
The bigger difference is that Alibaba Cloud itself is a cloud manufacturer, and it still has the Feitianyun operating system.
This means that, unlike hardware manufacturers such as NVIDIA and Intel, Alibaba Cloud has a more personal experience of the problems faced at various stages of cloud computing technology development.
As mentioned above, the CIPU is a processor specially designed for Feitian systems. That is to say, from the beginning of its design, it was designed to fit the pain points of the cloud computing industry and combine the characteristics of the Feitian system.
Such integration of software and hardware, on the one hand, can provide both high performance through hardware and flexibility through software.
On the other hand, the problem of adaptability is avoided from the beginning, and stronger performance, lower price, and higher stability can be achieved by means of 1+1>2.
Self-developed cloud computing
It seems that the first to enjoy this wave of technological development dividends are cloud users——
Cloud computing can achieve higher cost performance.
As the cloud manufacturer behind the CIPU, Alibaba Cloud’s technology release is also an important part of the localization of cloud computing technology in the new stage.
Take Alibaba Cloud itself as an example:
In 2009, Alibaba Cloud’s self-developed cloud computing operating system Feitian was born. The extreme concurrent scenarios that everyone talks about, such as Double 11 and 12306 Spring Festival ticket purchases, all run on this system.
In 2017, in order to solve the problem of server virtualization performance loss, Alibaba Cloud independently developed the Shenlong architecture (that is, the Shenlong card mentioned above). By transferring virtualization to dedicated hardware for acceleration, it achieved “zero loss” in performance. .
In terms of cloud storage technology, Alibaba Cloud’s self-developed Pangu distributed storage system has promoted the development of the ZNSSSD international standard for data centers. NVMe2.0, jointly proposed with Western Digital (WD), is currently the most advanced distributed storage system in the cloud computing industry with a deep integration of software and hardware.
Last year, Alibaba Cloud also released the first CPU Yitian 710, setting a new performance record for Arm server chips.
……
From network to storage, from software to hardware, through 13 years of technology accumulation and independent research and development, Alibaba Cloud, as a representative of domestic cloud manufacturers, is making more and more voices on the world cloud computing stage.
The launch of CIPU means that this effort in technology autonomy may have gone a step further:
Try to break the traditional development path defined by overseas cloud manufacturers and hardware manufacturers, and blaze a new path of your own.
Whenever technology develops to a new stage of upgrading, the competition around the right to speak is often exciting, and its impact is even greater than the technology circle itself.
5G is like this, and so is cloud technology.
The good show may have just begun.
Hashtags: Alibaba Cloud Alibaba CPU Processor