Summer in Moscow this year was, to be honest, not very good. It started too early and quickly, not everyone had time to react to it, and it ended already at the end of June. Therefore, when Huawei offered me to go to China, to the city of Chengdu, where their R&D center is located, after looking at the weather forecast at +34 degrees in the shade, I immediately agreed. Still, my age is not the same, and I already need to warm my bones a little. But I would like to note that it was possible to warm not only the bones, but also the insides, because the province of Sichuan, in which, in fact, Chengdu is located, is famous for its love for spicy food. But still, this blog is not about travel, so let's get back to the main goal of our trip - a new line of storage systems - Huawei Dorado V6. This article will wave you a little from the past, because. it was written before the official announcement, but published only after the release. And so, today we will take a closer look at everything interesting and tasty that Huawei has prepared for us.
The new line will have 5 models. All models except 3000V6 can be in two versions - SAS and NVMe. The choice depends on the interface of the disks that you can use in this system, Back-End ports and the number of disk drives that you can install in the system. For NVMe, Palm-sized SSDs are used, which are thinner than classic 2.5" SAS SSDs and can be installed up to 36 pieces. The new line is All Flash and there are no disk configurations.
Palm NVMe SSD
In my opinion, the Dorado 8000 and 18000 look like the most interesting models. Huawei positions them as High-end systems, and thanks to Huawei's pricing policy, it opposes these Mid-range models to the competitors' segment. It is on these models that I will concentrate today in my review. I note right away that due to their design features, junior two-controller systems have a slightly different architecture, different from Dorado 8000 and 18000, so not everything that I will talk about today is applicable to junior models.
One of the main features of the new systems is the use of several proprietary chips, each of which allows you to distribute the logical load from the controller's central processor and add functionality to various components.
The Kunpeng 920 processors, developed on ARM technologies and manufactured by Huawei independently, became the heart of the new systems. Depending on the model, the number of cores, their frequency and the number of installed processors in each controller vary:
Huawei Dorado V6 8000 - 2CPU, 64 core
Huawei Dorado V6 18000 - 4CPU, 48 core
Huawei developed this processor on the ARM architecture, and as far as I know, initially planned to install it only in the older Dorado 8000 and 18000 models, as it already did with some V5 models, but the sanctions made their own adjustments to this idea. Of course, ARM also talked about refusing to cooperate with Huawei during the imposition of sanctions, but here the situation is different than with Intel. Huawei manufactures these chips on its own, and no sanctions can stop this process. Breaking relations with ARM only threatens to lose access to new developments. As for performance - here it will be possible to judge only after conducting independent tests. Although I saw how 18000M IOPS was removed from the Dorado 1 system without any problems, until I repeat this with my own hands in my rack, I will not believe it. But there is really a lot of power in the controllers. The older models are equipped with 4 controllers, each of which has 4 processors, which gives a total of 768 cores.
But I will talk about the cores even later, when we look at the architecture of new systems, but for now let's get back to one more chip installed in the system. The chip looks like an extremely interesting solution (as far as I understand, the younger brother of the Ascend 910, which was recently introduced to the public). Its task is to analyze the data blocks arriving at the system in order to increase the Read hit ratio. While it is difficult to say how he will show himself in the work, because. today it works only according to a given template and does not have the ability to learn in an intelligent mode. The appearance of an intelligent mode is promised in future firmware, most likely at the beginning of next year.
Let's move on to architecture. Huawei has continued to develop its own Smart Matrix technology, which implements a full mesh approach to connecting components. But if in V5 it was only for access from controllers to disks, now all controllers have access to all ports on both the Back-End and Front-End.
Thanks to the new microservice architecture, this also allows load balancing between all controllers, even if there is only one lun. The OS for this line of arrays was designed from the ground up, not just optimized for Flash drives. Due to the fact that all our controllers have access to the same ports, in the event of a controller failure or reboot, the host does not lose a single path to the storage system, and path switching is carried out at the storage system level. However, using UltraPath on the host is not strictly necessary. Another "savings" in the installation of the system is a smaller number of required links. And if with the “classic” approach for 4 controllers we will need 8 links from 2 factories, then in the case of Huawei, even 2 will be enough (I’m not talking about the sufficiency of the bandwidth of one link now).
As in the previous version, a mirrored global cache is used. This allows you to lose up to two controllers at the same time or three controllers in series without affecting availability. But it is worth noting that we did not see full load balancing between the remaining 3 controllers in the event of a failure of one at the demo stand. The load of the failed controller was completely taken over by one of the remaining ones. It is possible that for this it is necessary to let the system work for a longer time in this configuration. In any case, on my own tests, I will check this in more detail.
Huawei positions new systems as End-to-End NVMe systems, but to date, NVMeOF is not yet supported on the front end, only FC, iSCSI or NFS. At the end of this or the beginning of the next, like other chips, we are promised RoCE support.
Shelves are connected to controllers in the same way using RoCE, and there is one drawback associated with this - the lack of a “loop” connection of shelves, as was the case with SAS. In my opinion, while this is a rather big disadvantage if you are planning a fairly large system. The fact is that all the shelves are connected in series, and the failure of one of the shelves entails the complete inaccessibility of all the others following it. In this case, to ensure fault tolerance, we will have to connect all the shelves to the controllers, which entails an increase in the required number of backend ports in the system.
And one more thing worth mentioning is non-disruptive update (NDU). As I said above, Huawei has implemented a container approach in the OS for the new Dorado line, this allows you to update and restart services without having to completely reboot the controller. Here it is worth mentioning right away that some updates will contain kernel updates, and in this case, the classic controller reboot will sometimes still be required during the update, but not always. This will reduce the level of impact of this operation on the productive system.
In our arsenal, the vast majority of arrays are from NetApp. Therefore, I think it would be quite logical if I make a small comparison with systems that I have to work with quite a lot. This is not an attempt to determine who is better and who is worse, or whose architecture is more advantageous. I will try soberly and without fanaticism to compare two different approaches to solving the same problem from different vendors. Yes, of course, in this case we will consider Huawei systems in “theory” and I will also separately note those moments that are only planned to be implemented in future firmware versions. What advantages do I see at the moment:
- Number of NVMe drives supported. NetApp currently has 288, Huawei has 1600-6400 depending on the model. At the same time, Huawei's Max usable capacity is 32PBe, like NetApp systems (to be more precise, they have 31.64PBe). And this despite the fact that drives of the same volume are supported (up to 15Tb). Huawei explains this fact as follows - they did not have the opportunity to assemble a larger booth. In theory, they have no volume limit, but so far they simply could not test this fact. But here it is worth noting that the capabilities of flash drives today are very high, and in the case of NVMe systems, we are faced with the fact that 24 drives are enough to utilize the top 2-controller system. Accordingly, a further increase in the number of disks in the system will not only not give a performance increase, but will also have a bad effect on the IOPS / Tb ratio. Of course, it is worth looking at how many drives can be pulled out by 4-controller systems 8000 and 16000, because. while the capabilities and potential of the Kunpeng 920 are still not completely clear.
- Lun's ownership of NetApp systems. Those. operations with the moon can only be performed by one controller, while the second one only passes IO through itself. Huawei systems, on the contrary, do not have any owners, and operations with data blocks (compression, deduplication) can be performed by any of the controllers, as well as written to disks.
- No drop in ports when one of the controllers fails. For some, this moment looks extremely critical. The bottom line is that switching within the storage system should be faster than from the host side. And if in the case of the same NetApp, in practice, we detected a freeze of about 5 seconds when pulling out the controller and switching paths, then switching to Huawei still needs to be practiced.
- No need to restart the controller when updating. This especially began to worry me with the fairly frequent release of new versions and firmware branches for NetApps. Yes, some updates for Huawei will still require a restart, but not all.
- 4 Huawei controllers for the price of XNUMX NetApp controllers. As I said above, thanks to Huawei's pricing policy, it can compete with Mid-range with its Hight-end models.
- The presence of additional chips in the shelf controllers and port cards, which are potentially intended to increase the efficiency of the system.
Cons and concerns in general:
- Direct connection of shelves to controllers or the need for a large number of back-end ports to connect all shelves to controllers.
- ARM architecture and the presence of a large number of chips - how efficient it will work, and whether it will be enough performance.
Most of the fears and fears will be able to dispel the hand-tested new line. I hope that soon after the release they will already appear in Moscow and there will be enough of them to quickly get one for their own tests. So far, we can say that, in general, the company's approach looks interesting, and the new line looks very good compared to competitors. a lot of questions are raised by the final implementation, tk. many things we will see only at the end of the year, and maybe only in 2020.
Source: habr.com