The word "cloud native" is very common. At the beginning, the most common phrase was "cloud native architecture". Later, many technical categories were added with the word "cloud native" in front of it, such as cloud native storage, cloud native database, cloud native network...
"cloud native" has no clear definition, but it has very obvious characteristics: it refers to a type of technology and practice that was born from public cloud or carried forward on public cloud.
cloud native processor is a name that has only been known in recent years. At first, I thought it was related to container cloud native. In fact, the concept of cloud-native processor was first proposed by Ampere, when the world's first cloud-native processor, Ampere Altra, was released in March 2020.
Later, I heard people from Amazon Cloud Technology introduce Graviton, saying that this is a cloud native processor. Recently, people from Ampere introduced Ampere Altra, and then I realized that I had a little understanding of the concept of cloud native processors, and gradually recognized the saying that this type of Arm processor is more suitable for public clouds.
Previously, I heard from people from Amazon Cloud Technology that Graviton2 was introduced to. At that time, the most impressive thing was not only the cost-effectiveness advantage of x86, but Graviton's single physical core provides a unique design with a single thread. A single physical core has a unique cache, which avoids the cores competing for cache and affects performance, making performance and latency performance more stable.
Ampere 80 core Ampere Alta and 128 core Ampere Alta Max can also achieve significant cost reduction and efficiency improvement. On the Tencent Cloud SR1 instance based on Ampere Altra, the key workloads in three web applications, NGINX, Redish, and MySQL, were tested. The cost-effectiveness shown is an average improvement of more than 50% compared with the S6 cloud instance based on x86 architecture (data source).
In addition, due to more Graviton cores than 64 cores, Ampere has more clearly demonstrated the advantages of multi-core. As the number of instances of workload (i.e., the number of running threads ) increases, the performance also shows a nearly linear increase, and the main frequency can reach more than 3.0GHz. When x86 grows to a certain level, the increase in the number of cores is very limited.
This is a picture shared by Sun Kaiben, product marketing director of Ampere Computing, on 2022 OCP China. The test runs the Encoding encoding workload. The horizontal coordinate is the number of concurrent encoding tasks, which represents the number of threads used, and the vertical coordinate is the performance data. As the number of tasks increases, there are quite a difference in the trend of performance growth.
Another picture shows the performance predictability of Ampere Altra Max, which can also be said to be the performance stability. The test above
is running Redis workloads. While running Redis loads, other workloads are inserted periodically. As the load is added and completed, the performance changes of the main workload are observed. As shown, the performance of Redis running on the Ampere Altra Max is almost unchanged, while the performance jitter of the x86 is very obvious.
Sun Kaiben said in his speech that there are two main reasons: one is the single core and single thread design, and the other is due to excellent power consumption control.
SPEC test: The actual power consumption of the Ampere processor is lower than TDP, so that the main frequency can be stabilized
Generally speaking, the design of single core and single thread reduces mutual interference between cores, which not only helps to stabilize performance, but also helps to increase performance with the increase of core count. At the same time, with the advantages of power consumption control, Ampere not only allows all running cores to maintain a consistent maximum main frequency, thereby improving stability, but also makes it possible to have more cores heaps.
power consumption is an important consideration in chip design. The main frequency of the Ampere processor is relatively low and the fluctuation range is also very small, which is convenient for controlling power consumption. In order to maintain stable performance, it can even remain at a certain main frequency. There is no need to worry about hitting the power consumption wall and reducing the frequency and causing performance jitter.
Of course, x86 can also turn off hyperthreaded technology and lock the frequency to a lower level to provide more stable performance.Although the single core performance of x86 is stronger at this time, due to the small number of cores, the overall performance is still unknown. Moreover, this approach is like a waste of martial arts that have been practicing for many years. At present, there should be only a few users who want to use x86 processors like this. Moreover, for cloud service providers, turning off hyperthreading means that the number of copies available on the same platform can be halved, and the economic benefits will be greatly reduced.
In short, single core provides single threading and power consumption advantages, Ampere has more stable performance and more linear performance growth performance, and this is exactly what public cloud manufacturers need:
On the one hand, public cloud hopes to use innovative technologies to reduce costs and increase efficiency. The core advantage of
Ampere greatly improves the computing density of cloud data centers and saves the space cost of data centers. At the same time, being able to use a single device can provide more instances and serve more users. Ultimately, the results of cost reduction and efficiency increase will eventually benefit users. Currently, by using cloud instances based on Ampere Altra (Ali Cloud c6r), the domestic observable solution provider Observation Cloud has also achieved a 40% cost-effective improvement. This can also be referred to Graviton in AWS.
On the other hand, the public cloud hopes to provide users with a better user experience.
Many users are worried that the "noisy neighbor" problem affects performance, and the performance of the Ampere solution can maintain a stable performance at all times, allowing cloud service providers to provide consistent and stable services and avoid performance jitter affecting business.
or above is the main reason why cloud manufacturers choose Arm servers.
OCP China Day 2022 event theme has the keyword "green". As a global power user, the data center's work in green and low carbon is also very important. The advantages of Arm in low power consumption will also help data centers achieve low carbon goals, which will also be a major driving factor in the adoption of Arm servers in data centers.
Although x86 architecture will still be mainstream for a long time in the future, there will be more and more users of Arm servers. Today, Arm's application ecosystem has made a lot of progress.
On the one hand, AWS's self-developed Graviton is a banner of the Arm camp, which has made more people pay attention to and even started using Arm servers. Arm chip manufacturers like Ampere can share the Arm application ecosystem without barriers.
On the other hand, Ampere's third-party processor is used by major public cloud manufacturers such as Oracle, Google Cloud, Tencent Cloud , Microsoft Azure, Ali Cloud , etc. The driving force of these leading manufacturers will not only accelerate the development of the ecosystem, but also allow the market share of Arm server CPUs to grow rapidly. According to data from research company Omdia, global server shipments in the second quarter of this year were 3.4 million units, of which the Arm architecture reached a record 7.1%, and the share of increased by 48% year-on-year.
In addition, server manufacturers such as Inspur , Ultramicro , HPE, Weiying, Gigabyte have also launched server hardware, which not only serves the super-large data center market, but also serves corporate users.
Ampere official website lists six major scenarios supported by Ampere, including: various common machine learning frameworks, supporting Arm's old business (cloud mobile phones, IoT, autonomous driving), big data analysis workloads, cloud-native application loads, cloud storage and web services.
Among them, there are relatively few cloud storage parts, and only Ceph is currently displayed. The cloud native application load part is very rich, and there are many listed on the official website. You can see 4943 images on DockerHub, which is about one-third of that of x86, and you can have all kinds of commonly used images that you can see.
In May this year, Ampere announced plans for the next generation processor AmpereOne™, which will adopt the 5nm process process , supports PCIe 5.0 and DDR5, and will use customized cores, thus giving more flexible design space. Ampere also said that it will maintain the pace of annual product updates in the future.
ProLiant RL300 released by HPE not long ago is a single-channel server, but at the OCP China Day 2022 event site, two servers were placed on Ampere's booth, both of which are dual-channel servers, and dual-channels can provide at least 160 cores.
On the left is Inspur NF5080R6, and on the right is an Mt. Jade
One is Inspur dual-channel server NF5080R6 using Ampere processor ("R" probably means Arm). The other is the open source Mt.Jade server, which is an open source server contributed by Ampere to OCP. Anyone can download design drawings for manufacturing and use.
Some experts said that this is due to the procurement habits of Chinese users and may improve in the future.
