Today, there are many players in the SDS field, and there are many technical schools. There are open source and closed sources. There are symmetric structures and asymmetric structures. There are SDS, also called HCI (Hyper-Converged Infrastructure), there is also called ServerSAN, cloud storage. , dazzling, incomprehensible. It should be said that SDS is still in the early stages of development, and the corresponding technologies and test standards are still not perfect, and the depth of customers' understanding of their technologies is not enough. The combination of these two factors will make it difficult for customers to choose SDS. Three key indicators of three types of SDS application scenarios SDS is divided into block storage, file storage, and object storage according to the provided data organization. According to the delivery model, it is divided into two types: independent storage form and HCI form (that is, integrating SDS, calculation and network in one system). This article only takes SDS's block storage category as an example, how to choose a suitable SDS from the perspective of enterprise users. The main application scenarios of SDS block storage are similar to those of traditional SAN devices. They are mainly used in scenarios that require the provision of bare disks (block devices), such as server virtualization, desktop virtualization, and databases. Of course, the block devices provided by the SDS block storage can also be formatted in the local file system and store some unstructured data, but this is not the main application scenario. When suggesting customers to choose SDS, first of all, we need to ask ourselves two questions: 1. What are the main application scenarios? 2. What are the top three technical indicators most concerned about? It is also necessary to know at what stage SDS is being developed (at present, SDS is still in the initial stage of socialism) and to bring expectations and reality closer. Some of the overly high demands are still not available at this stage. About SDS technology evaluation usually includes the following seven dimensions: reliability, stability, scalability, functionality, performance, ease of use, and compatibility. The first three technical indicator dimensions for SDS blocks are: The three major application scenarios put reliability and stability in the top two. The reason for this arrangement is that storage is the cornerstone of IT systems. The so-called reliability is that the local failure will not lead to data loss and business interruption, which should be ranked first; stability means that the performance will not be greatly shaken due to local failure, affecting the business response, such as A small movie midway old card who can stand it. Speaking of this, some people will ask, these two points are the necessary and default requirements for the traditional array? I would like to say that traditional arrays are expensive, dedicated hardware heaps, and the probability of hardware failures is low. Most SDSs are deployed on cheap standard x86 servers. The probability of hardware failure on x86 servers is much higher; and SDS is much higher. It is a distributed system. To manage dozens or even tens of thousands of x86 servers, the probability of such a hardware failure will increase by an order of magnitude. How to ensure the reliability and stability of a large-scale distributed system is also the most difficult problem for SDS. Amazon's S3 system is made up of millions of servers. There are many devices going offline and on-line every minute. The reliability and stability of such a large-scale cluster can be ensured. Should few manufacturers have such technology? Also, don't forget that SDS is still in the "primary stage of socialism." It's not that a single manufacturer can guarantee the reliability and stability of the SDS system for thousands of servers. The VSI and VDI environments are usually very large, and the speed of scale growth is also faster, so scalability is placed third. Performance is a very important indicator for database applications, so it ranks third. Seeing may not be true, those "dirty" in the test Finally I was wondering what kind of SDS I wanted. Suddenly I remembered a big event: "Who's got my product?" One of the most straightforward ways is to take the test as an example (the test is still quite troublesome, prepare the test environment, write the test specification, view the test results...). Maybe there is a lazy way, that is, "Do you have a similar case to my needs? I asked the company to use the same kind of ah," but hearing is imaginary and the word is not always practical. . In contrast, testing is a more reassuring method, but if you are not familiar with the water of SDS, there are also some pits you can't find out. Well, the following paragraph is to help customers judge the merits of key technical indicators from the perspective of testing and technology architecture. Reliability is mainly reflected in two aspects on the SDS. When the disk or node in the cluster fails, will the data be lost? Will business be interrupted? What is the duration of the interruption? There are two indicators that need attention: 1. Fault tolerance, 2. Fault recovery time. Let us talk about the tolerance factor. Because the mainstream SDSs all use the copy technology. For example, if the three copies are lost, there should be 2 data, and the data will not be lost. That is, the fault tolerance is 1/3 node. But if more than 1/3 of the nodes are down at the same time, can the cluster still work? This is not necessarily all manufacturers can do. Many fault-tolerant domains of SDS are pre-configured. Taking 3 replicas and 9 nodes as an example, three fault domain A\B\C are usually configured and each fault domain has 3 nodes. Each fault domain saves independent replica data. For example, when three machines with a fault-tolerant domain A all fail, two or two copies exist, data will not be lost, and the business will run as usual. This is commonly referred to as 1/3 node failover. Most manufacturers can do this, but if the B-domain or C-domain is also machine down machine? Most of the time, the two copies will be lost and the system will be abnormal. This indicator of failure recovery time is also critical. When a node is down in the system, the time of failure recovery is better as soon as possible. The highest requirement is that the failure recovery time is equal to 0. To achieve this goal is not easy, because when the node in the SDS system fails, to restore this failure, you need to do a lot of things: the first step, find this failure; the second step, select a node to replace the failure node, and proceed Data reconstruction; The third step is to refresh the address space so that the client can obtain the change in the data location. Each step takes time, especially for large-scale clusters, and it is difficult to control the recovery time to a small extent. There are not many SDS manufacturers claiming that the fault recovery time is zero. So the magnitude of fault recovery time is a very important factor to measure a SDS reliability level. Users can set corresponding requirements according to the sensitivity of front-end services to fault recovery time. For SDS, the main concerns of its stability are: When the disk/node failure occurs in the system and data is restored due to data migration, the front-end performance is stable. As mentioned in the previous reliability paragraph, there are three steps to SDS failure recovery, and poor handling at each step can affect performance. In particular, data reconstruction and re-addressing have a great impact on performance stability. For example, the well-known open source product Ceph, more than one customer reported that in the event of node downtime in the system, performance degradation and fluctuations are very severe and have a great impact on the business. The only reason why this problem first arises is related to its way of metadata management and addressing (Crush algorithm). That is, when the node is turbulent, the metadata (in fact, the list of resources saved inside the ceph) changes, which causes the address of the data to occur. Changes eventually lead to a large number of unnecessary data migrations. The larger the scale, the more obvious this problem is exposed. Second, it uses an entire disk copy for data migration, and there will be congestion in the network bandwidth caused by invalid migration. There is also a pit that reveals that some SDS systems may not reconstruct data when they are pulling or detaching nodes, and reconstruct the disk or node when they are brought online or restored. Therefore, it is best to observe the performance of the disk or the node after the stability test. It is even recommended to test its stability with frequent node up/down, because the node jitter will still occur from time to time. Extensibility is hard to test because it is impossible for you to prepare several hundred or thousands of server environments unless you are a local tyrant. What to do? No way, look at the architecture. The mainstream SDS on the market is divided into two major categories: an asymmetric architecture with a central metadata management node and a fully symmetrical architecture without a central management node. The former has limited scalability compared to the latter. The simple understanding is that there is a commanding officer in an "asymmetrical framework" and that there is a limited number of people who direct the functional management. There is no commander in the "full symmetry framework". It is self-conscious. It is like a lot of people working. If one is sick, there is no need to ask the leader for leave. There will be another person who will automatically take over his job. As an example, VSAN is a central management node. It officially claims that the single cluster supports 64 nodes. The ZettaStor network of Pengyun Cloud has no central node and can support 10,000 nodes. Therefore, as can be seen from the specific examples, the architecture determines its scalability. Currently, flash memory technology has developed rapidly and its performance is several orders of magnitude higher than that of conventional magnetic media drives. It is precisely because the development of flash memory technology has also given SDS a basis for PK-based traditional arrays in terms of performance. If you are rich, you can use SSD as the main memory. However, SSDs are still expensive, and the cost-effective way is to use SSDs as caches. Usually, a storage node is equipped with a few hundred GB-1TB SSDs as caches. The SSD cache's contribution to performance will perform very well when the amount of small data is small. Once the amount of data is large enough and the SSD is penetrated, then the performance of the disk (write to the hard disk persistence layer) will be demonstrated. If your system has a small amount of data and the capacity of the SSD cache is sufficient to support your business peaks, you can pay more attention to the performance of SDS cache acceleration. If your system has a large amount of data, then SSD will be penetrated for a long time, then you will focus on the performance of a falling disk. At present, the SDS manufacturers use the SSD cache to reduce the utilization efficiency, and there are not many unique algorithms. Instead, this is a part of the process because of different technical routes and different performances. Some have almost played the performance of the original disk, and even low; some use some IO algorithm to improve the performance of ordinary disk several times, like Pengyun network's ZettaStor used a "randomized to semi-order The IO optimization algorithm improves the speed of random IO raw disks by 3-5 times. Peng Yun network can do this kind of IO optimization because ZettaStor is completely self-developed. Manufacturers adopting open-source solutions will have difficulty realizing this, and this is to move to their core file system layer. Some manufacturers will test with small volumes and small data during performance testing. It seems that the IOPS will be very high, but the real environment will not be such a small load, so once more large volumes are created, it will take a long time. The data volume performance test, when the SSD was written and penetrating, the performance plummeted and the Phoenix turned blackbird. Do not remember the big coffee who said a word, "Do not talk about delays in performance tests are rogue." Looking at the delay of a system is not small, one is measured, but also from the framework can see some clues. It is to see if its IO path is short enough. For example, writing bare disk directly is much shorter than the IO path of the bare disk after file system layer conversion. This is why Ceph wants to reduce the delay. As we all know, Ceph's block storage does not directly access bare disks, but instead converts raw disks into block devices for application through the file system. Having seen these questions, is it that you are deterred from SDS? If this is the case, you will spill the child and the dirty water together. Just like the reform and opening up, the sun and the flies will always come in together. The good and the bad are different. This is a normal phenomenon in the current SDS market. The reason why Peng Yun opened up some scars is because I believe that real gold is not afraid of fire! Users can not choose Pengyun network, but we must strive to choose the right! This is what this article is looking for!
Our Surface Mount Box could include the CAT5E CAT6 CAT6A RJ45 Connectors
The surface mount box is an ideal for wall box and wire installation, we design many junction box, back mount box, blank mount box, loaded surface mount box.
Mateiral: ABS, PBT UL94V-0
1~8 ports and colors supply for your choice, they are an ideal for low voltage application for your security, data, networking solutions and telecom projects.
CAT5E Surface Mount Box,Surface Mount Box,Adapter Inline Coupler,single port junction surface mount box NINGBO UONICORE ELECTRONICS CO., LTD , https://www.uonicore.com