Having taken a bird's eye view of all the modern Huawei Enterprise solutions presented in 2020, we move on to more precise and detailed stories about individual ideas and products that can serve as the basis for digital transformation of both large enterprises and government agencies. Today - about what concepts and technologies Huawei proposes to build data centers on.
In the era of the "connected world", the tasks of storing and processing data require new approaches at all stages of the life cycle of data centers. They must simultaneously become both simpler and “smarter” in order to cope with the role entrusted to them as the central elements of the infrastructure of the global digital economy.
In 2018, humanity stored 33 zettabytes of information, but by 2025, its total volume should increase by more than five times. Three decades of experience in the development of ICT infrastructures have enabled Huawei to prepare well for the growing “data tsunami” and offer its partners and customers the concept of an intelligent data center, which includes all stages of its construction, operation and maintenance. The elements of this concept are united under the general name HiDC.
Digitize it
A fresh anecdote is circulating on the Web: who accelerated the digital transformation of your company the most - CEO, CTO, board of directors? Coronavirus epidemic! Only the lazy one does not hold webinars, does not write articles, does not tell how and what to do. But these are all reactive actions. Some were pre-prepared.
Not for the sake of bragging - for objective reasons, as an example, we will cite our company, in which digital transformation was massively initiated several years ago. At the moment, we are able to transfer almost all employees to work from home without any loss of efficiency. The story of the hospital built in the city of Wuhan in ten days is indicative. There, digital transformation manifested itself in the fact that all IT systems were deployed in three days. So digital transformation is not about “when” and “why”, but about “how”.
Architectural approach instead of spontaneous development
What are the main problems we face when we start building a system? Until now, all our customers have been working in the mode of combining business tasks with applied services and IT solutions. It is quite difficult to get a general idea of the functioning of such a complex if it was created simply by adding various blocks. And in order to build a system as a single organism, first of all, an architectural approach is needed. We embodied it in the ideology of our HiDC solution.
Maximum value and minimum value
The entire structure of HiDC is two main slices. The first is what you are used to seeing from Huawei - the classic infrastructure. The elements of the second slice are easiest to combine with the term "smart data".
Why is this needed? Today, many companies accumulate colossal volumes of information, often scattered or accessible through various kinds of “pads”. Yes, take at least ordinary databases. Ask your DBAs how these databases fit together and how to use the information from them in BI systems to make business decisions. Surprisingly, often the databases are very loosely connected to each other and function as separate "islands". Therefore, first of all, we thought about what architectural approaches this problem could be eliminated.
HiDC Architecture Design Principles
Consider the basic design principles of HiDC. First of all, this will be useful not for specialists in a particular area, but for solution architects who can capture the entire panorama with a glance.
The most common are the converged networks block and the data management block. And already here a concept arises that solution architects rarely think about - data lifecycle management. From classical databases, it has migrated to many other systems, including cloud and edge computing.
There are more and more edge computing. The most obvious example of their application is a car with an autopilot, which it is desirable to control from a single platform. In addition, there is a trend towards "green" technologies - more energy efficient, causing minimal damage to the environment. Both can be achieved by switching to intellectual resources (more on them later).
It's great when we have all six blocks of the HiDC structure at our disposal. True, often customers work in a previously created environment. However, the use of even one block from the diagram above can bear fruit. And if you add a second, a third, and so on, a synergistic effect will begin to appear. The combination of networking and distributed storage alone will give you better performance and lower latency. The block approach allows us to develop not chaotically, as often happens in the industry, but using an integrated architectural approach. Well, the openness of the blocks themselves provides freedom in choosing the optimal solution.
Time of Converged Networks
Recently, in the global and Russian markets, we are increasingly zealously promoting the concept of converged networks. Already today, our customers use converged solutions based on RoCEv2 (RDMA over Converged Ethernet v2) to build distributed software-defined storage systems. The main advantage of this approach is its openness and the absence of the need to create an indefinite number of disparate networks.
Why wasn't this done before? Recall that the Ethernet standard was developed in 1969. For half a century, many problems have accumulated in it, but Huawei has learned how to solve them. Now, thanks to a number of additional steps, we can use Ethernet for mission-critical applications, high-load solutions, and more.
From DCN to DCI
The next important trend is the synergistic effect from the introduction of DCI (Data Center Interconnect). In Russia, unlike China, something similar can only be found with telecom operators. When customers consider data center networking solutions, they usually don't pay enough attention to the deep integration of optical networks and classic IP solutions within a single point of presence. They use familiar solutions that work on the IP layer, which is enough for them.
What is DCI for then? Imagine that the DWDM site administrator and the network administrator act independently. At some point, a failure that occurs with any of them can seriously reduce your fault tolerance. And if we use the principle of synergy, IP routing is carried out taking into account what happens on the optical network. The use of such an intelligent service significantly increases the number of nines in the availability level of the entire system.
Another major advantage of our DCI is a large performance margin. Summing up the possibilities of the C and L ranges, you can get about 220 lambdas. Such a reserve is unlikely to be quickly exhausted even by a large corporate customer, given that our current solution allows you to transfer up to 400 Gb / s through each lambda. In the future, on the same equipment, it will be possible to achieve 800 Gb / s.
Additional convenience is provided by the overall manageability we provide through classically exposed interfaces. Through NETCONF, not only switches, but also optical multiplexing devices are managed, which allows achieving convergence at all levels and perceiving the system as an intellectual resource, and not a “set of boxes”.
Edge Computing Is Increasingly Important
Many have heard about Edge Computing. And those involved in cloud and classic data centers should keep in mind that we have recently seen a serious shift towards edge computing.
What caused it? Let's look at common implementation models. Now there is a lot of talk about "smart cities", "smart houses", etc. This concept allows the developer to create added value and increase the price of the property. "Smart home" recognizes its resident, lets him in and out, provides him with some services. According to statistics, such services add about 10–15% to the price of apartments and, in general, can push the development of new business models. Also, it has already been said about the concepts of autopilots. Soon, the development of 5G and Wi-Fi 6 technologies will allow for extremely low latency data transfer between edge-computing smart homes, cars and the main data center. This means that it will be possible to perform a much larger number of operations related to serious data processing. To solve such problems, in particular, it is possible to use neural processors that are already being shipped to Russia.
The prospects of the trend just outlined are undeniable. Imagine, for example, an intelligent city traffic management system that can switch traffic lights, regulate traffic on specific streets, or even take adequate measures during emergencies.
Now let's turn to the resources with which we provide the implementation of the HiDC concept.
Calculations
When we need to implement a standard computing system, x86 processors, of course, find application in it. But as soon as the need for customization arises, it is time to think about more diverse solutions.
So, for example, ARM processors, due to the large number of cores, are great for applications with a high degree of parallelism. Multithreading gives a performance gain of about 30%.
If low latency is critical to us, field-programmable logic integrated circuits (FPGA) FPGA come to the fore.
Neural processors are primarily needed for solving machine learning problems. If for a specific implementation we need 16 racks with 8 servers each, stuffed with neuroprocessors, then a solution of the same level based on the x86 architecture would require (!) about 128 racks. As you can see, a wide variety of calculation types forces you to carefully choose hardware platforms.
Data Storage
For the second year now, Huawei has been calling on partners, customers, and industry colleagues to build storage systems in accordance with the Flash Only principle. And most of our customers only use mechanical spindle drives for legacy applications or rarely used archive data.
Flash systems are also evolving. Storage Class Memory (SCM) systems such as Intel Optane are entering the market. Interesting developments are demonstrated by Chinese and Japanese manufacturers. At the moment, SCM is superior to all other solutions in terms of processing class. So far, only the high cost does not allow their use everywhere.
At the same time, we see that the quality of storage needs to be improved not only on the conditional backend, but also on the frontend. Now, de facto, in new implementations, we, as a rule, offer and use mechanisms for direct memory access via Ethernet, but we see a request from customers and therefore, towards the end of the year, we will start using NVMe over Fabrics more often. Moreover, end-to-end, in order to provide a common architecture, which, of course, must be high-performance and resistant to controller failure.
OceanStor Dorado storage is one of our flagship products. Internal tests have shown that it delivers 20M IOPS performance while maintaining XNUMX out of XNUMX controller failures.
Why such power? Let's take a look at the current situation. For several months now, Chinese people have been spending significantly more time at home due to the lockdown. Internet traffic at this time increased by an average of 30%, and in some provinces even doubled. Consumption of various network services has grown. And at some point, the same banks began to experience a serious additional load, for which their storage systems were not ready.
It is clear that not everyone needs 20 million IOPS now. But what will happen tomorrow? Our intelligent systems maximize the potential of neural processors to ensure traffic compactness, deduplication, optimization and online data recovery.
Reference network
2020, as we mentioned in the previous article, will be the year of core networks for us. Many customers, especially application service providers (ASPs) and banks, are already thinking about how their applications will work specifically in terms of data center and inter-data center connectivity. This is where a new backbone network comes to the rescue. As an example, let's take the largest Chinese banks that have switched to simplified backbone systems that use not a dozen different protocols for communication between data centers, but, relatively speaking, a couple - OSPF and SRv6. Moreover, the organization receives the same set of services.
Intellectual resources
How to use data? Until recently, there was a fragmented system of diverse databases: Microsoft SQL, MySQL, Oracle, etc. To work with them, big data solutions were used that could combine this data, take it, work with it. All this created a high load on resources.
At the same time, there was no mechanism for performing operations with data upon the occurrence of some event. The solution was the development of data lifecycle management (DLM) principles.
Everyone has heard of data lakes. With the transition from data management to data governance, “digital lakes” began to rapidly “get smarter”. Including thanks to Huawei solutions. In the following materials, we will definitely tell you about the entire stack of software technologies we used. Now it is important to note that it was the use of "smart" data lifecycle management that allowed us to simplify the use of our network and servers, as well as learn how to build end-to-end architectures for a better understanding of the principles of working with data.
Engineering infrastructure of the data center
We will publish separate materials on engineering infrastructure, however, in the context of today's topic, we would like to mention those changes that are related to the HiDC concept.
For a long time, the use of lithium batteries in emergency and backup power systems (PSS) of data centers was banned due to their high fire hazard. Any mechanical damage or violation of the integrity of the battery could lead to its ignition and unpredictable consequences. In this regard, the SRPs were equipped with obsolete acid batteries, which have a low specific charge density with a large mass.
Huawei's new emergency and backup power systems use safe lithium iron phosphate (LFP) batteries with intelligent proactive management. With the same capacity, they take up three times less volume compared to acid batteries. Their life cycle is 10-15 years, which, among other things, reduces the burden they create on the environment. The SmartLi ecosystem's patented control system allows hybrid systems of old and new battery arrays to be used, and the switching system allows hot changes to the PSA structure while maintaining the redundancy function.
Smart operation
An important part of the operating principles of the HiDC infrastructure is the smart self-healing ideology. IN From our previous publications, we mentioned the intelligent O&M 1-3-5 platform, which is able not only to detect and analyze an undesirable event in the system, but also to offer the administrator several options for a fully automated solution to the problem.
The introspection function allows you to detect problems in about a minute. Three minutes are spent on analysis, and within five minutes proposals are formed to change the state of the system.
Let's say some operator error has led to the formation of a closed loop of processes that reduces the performance of a virtualization farm from 100% to 77%. The data center administrator receives a corresponding message on his dashboard, which contains a complete visualization of the problem, including the network diagram of the resources affected by the unwanted process. Then the administrator can start correcting the situation manually or use one of several automatic recovery scenarios offered to him.
The system knows about 75 such scenarios that can be implemented in less than ten minutes. Moreover, they cover 90% of the problems that occur in data centers. At this time, the engineer can calmly answer the calls of concerned customers, being sure that the service will be restored any minute.
New key products in HiDC
In addition to software products, this includes key solutions that operate at the infrastructure level. First of all, we need to mention the neural processors used in our Atlas family AI clusters, as well as NPU and GPU-based servers.
In addition, we can not fail to mention the Dorado again and its class-leading performance, which will last for years to come. This is especially true in the post-Soviet space, where, with rare exceptions, it is customary to update something only when it completely stops working. This explains the service life of individual storage systems, reaching ten years. Enormous performance is necessary for Dorado to provide high-quality services ten years from now.
Innovation in every element
When choosing specific infrastructure solutions, one should not forget about the architecture and scenarios for its further development. Disparate products from different vendors do not guarantee the expected synergies that solutions that are already optimized for joint use will provide.
The infrastructure must be based on the right technologies. The “correct” ones are, among other things, open, providing high throughput, and functioning stably under high loads. In relation to data centers, for example, a good ratio of total energy consumption to IT load is important. To achieve all these goals, you need to choose the environment and components. In modern conditions, this also means an increasing use of artificial intelligence.
According to our observations, among the strategic customers of Huawei, there are fewer and fewer who still do not use machine learning systems. Without ML, it will simply not be possible to monetize the accumulated data as much as possible.
The monetization system can be different: banks offer new targeted products, telecom operators provide individual services and ensure loyalty, government customers need high-quality data lifecycle management and a high level of interaction with other organizations. After all, data management models are no longer limited to setting up a firewall and ensuring network visibility of their databases.
From idea to operating data center
The construction of a standard data center, at best, takes from a year to a year and a half. Our production cycle allows us to do this much faster through the use of a group of solutions, united under the common name FusionDC 2.0. Engineering, high-level design, assembly of all elements of the IT load are carried out directly at the factory. In a short time, equipment is delivered by sea containers from China to Russia. As a result, the creation of a turnkey data center can be achieved in just four to five months.
The idea of a prefabricated cloud data center is also interesting because it is possible to develop a data center in stages, adding the necessary functional blocks to it. This approach is embedded in the very concept of HiDC.
In order not to turn the review material into a datasheet, for additional information on HiDC, we suggest going to . There you will find a description and examples of the implementation of those approaches, products and solutions that we talked about. There will be more materials, the higher your level of access to the site. If you have been assigned the status of "partner", you will be able to download HiDC roadmaps, technical presentations, videos.
We would venture to suggest that the majority of those reading this article have the competencies of network architects. They will certainly be interested in visiting our . There we talk in detail about how to build a network infrastructure according to the rules of Huawei Validated Design (HVD). Guidelines available for download will help you thoroughly understand how the company's solutions work. Just remember that without authorization you will have access to fewer materials.
***
Numerous webinars, held not only in the Russian-speaking segment, but also at the international level, will also help you navigate. On them we share both information about our products and our business practices. In particular, we talk about how Huawei, in the face of a break in many service chains, continues to ensure the continuous delivery of its products to different countries. Recently, for example, there was a case when newly manufactured data center equipment reached a Moscow customer in just three weeks.
List of webinars for April is available .
Source: habr.com