Shulou(Shulou.com)11/24 Report--

Recently, Tianyi Cloud and technology media InfoQ held an online technology sharing meeting with the theme of "New solution to Storage problems, revealing the extremely easy-to-use HBlock". Pan Yalian, Director of Storage Product Line of Tianyi Cloud International Business Division, and Wu Zhimin, R & D expert, participated in the first phase of sharing, with the theme of "Revolutionary Software-defined Storage HBlock, Technical Analysis behind extreme ease of use", sharing HBlock product concept, application scenario, technical architecture design and other content.

Tianyi Cloud launches revolutionary software-defined storage HBlock

IDC predicts that the amount of data in China will grow to 48.6ZB by 2025, which brings all-round challenges to the capacity, performance, operation and maintenance, cost and flexibility of enterprise storage systems. In addition, due to overconfiguration, data islands, redundant backup, low access speed, inefficient search, and disruptive transmission and migration, there is a problem of low utilization of storage resources in the computer room.

In order to solve the problem of high cost and low efficiency of storage brought by heterogeneous storage and data isolated island, Tianyi Cloud developed the first storage resource activation system-HBlock.

Overall view of HBlock

HBlock is a full-user-mode software-defined storage product, which can be completely decoupled from the underlying infrastructure and coexist with the existing business on the server. It can uniformly manage the general X86, ARM, Godson and other servers and the storage resources managed above into a high-performance virtual storage array, and provide distributed block storage services through the standard iSCSI protocol.

Efficient management and easy activation

Users can manage and use their storage resources comprehensively and efficiently through HBlock, which is realized through various functions of the management plane. From the management plane, HBlock is divided into four layers: access layer, verification layer, service layer and data layer.

HBlock management plane

When ★ accesses the access layer, users can use API, command line and management console to manage HBlock flexibly, intuitively and comprehensively.

★ is in the verification layer, and HBlock uses HTTPS, authentication and interface verification to ensure the security of access.

★ is in the service layer. Users can choose to create stand-alone or cluster servers according to their local hardware resources. Deployment can be completed in 3 minutes to create, delete, expand, count, monitor and other storage resources. If there are any problems during use, Tianyi Cloud engineer can assist remotely and connect to the system for troubleshooting. If privacy issues are taken into account, you can also send the operation log to Tianyi Cloud engineer for troubleshooting. Users can also set an e-mail address to receive the alarm information of the system in time.

★ in the data layer, HBlock unified management of users' data, through a unique technical architecture to achieve high performance.

The key technologies behind the extreme ease of use of HBlock

HBlock not only provides users with efficient, comprehensive and flexible management means, but also provides extremely high performance. Behind all this is the support of excellent technical architecture. The technical architecture of HBlock is divided into three layers from top to bottom: access layer, storage service layer and operating system and hardware layer.

HBlock technical architecture

At the top is the access layer, which provides block storage services through iSCSI, management services through API, command line and Web interface, and also supports OpenStack, CSI driver of K8S and other management methods.

In the middle is the storage service layer, and its core is the distributed lease module, which is completely self-developed by Tianyi Cloud, which does not require strong consistency of clock or frequency, so it can provide consistency guarantee for data service and management service.

The storage service layer also provides obsolete detection function. Before installing HBlock, users can use this tool to comprehensively test disk, network, CPU, memory and other performance dimensions, so that users can have a benchmark understanding of the performance of each component as soon as possible and identify potential performance weaknesses.

Another important module in the storage services layer is caching, which is the focus of HBlock high-performance optimization, which will be described in more detail later.

The storage service layer also includes HA, data redundancy, failure domain, data reconstruction and other modules, providing high availability and high reliability. IO scheduling, indexing, data balancing module, according to the intelligent scheduling algorithm, ensure that each node, each disk load balance, timely remove the components that cause performance bottlenecks.

At the bottom of the operating system and hardware layer, HBlock achieves true hardware heterogeneity. For example, for different sizes of disks, when allocating data, different weights are dynamically assigned according to disk capacity, load, performance, network delay and other information, and then different allocation methods are calculated. to ensure that each component performs the best in the cluster.

1. Multi-pattern matching of write cache

As a block storage product, the performance of HBlock is most directly reflected in the ability to read and write data. HBlock achieves high performance through the design of write cache and read cache. Next, let's take a closer look at the process of writing and reading data from HBlock.

First, take a look at the write data flow:

The client initiates a write request, and after the data is normally placed in the write cache, it continues to the write cache of the second slave node, and the success of the write will be reported to the client only after the write is successful. This can not only ensure that the data is retained on both nodes (avoiding a single point of failure), but also ensure that the data does not need to fall off the disk and end to the cache, and the IO path can be as short as possible, so the write latency is relatively low. This is the basic process of HBlock writing, also known as WriteBack.

The data is in the write cache and the disk is dropped according to different policies. There are quantitative policies, such as when dirty data reaches a certain proportion, fixed-time policies, such as dirty data reaching a certain amount of time, and active policies, such as when clients send sync commands on their own initiative, they will also fail. After the disk is dropped, the data is put into the read cache, and then the index is updated.

Some users may worry that although a single point of failure has been avoided, if the data has been written to memory and the write is successful, but before the disk is dropped, if the cluster as a whole fails, the data will also be lost.

In response to this situation, HBlock designed two other write cache modes: WriteThrough and WriteAround.

In WriteThrough mode, after the data reaches the write cache, it will continue to wait for the data to be dropped to the disk, and then report back to the client successfully after the data is actually on the disk.

WriteAround is very similar to WriteThrough, except that the last step is that the data will not enter the read cache after the disk is dropped.

These three different write caches have their own applicable scenarios. For example, WriteBack, its data ends in memory, so its read and write performance is relatively high. WriteThrough and WriteAround have to drop data every time, so its write performance is lower than that of WriteBack. For WriteAround, because the final data does not enter the read cache, it is more suitable for scenarios where there are more writes and less reads, and accordingly it takes up the least memory. Therefore, different configurations can be made according to different business scenarios.

The three caching modes are summarized as follows:

With regard to WriteThrough and WriteAround, because each write will drop the disk, which will significantly increase the number of system calls and, more seriously, the number of sync to the disk. Therefore, HBlock designs an adaptive write cache to dynamically adjust the timing of sync merging according to the concurrency of the client and the load performance of the disk, to balance the write request waiting time and the underlying disk pressure.

2. Optimization mechanism of read cache and prereader.

After you understand the data writing process, next, take a look at the data reading process:

When the read request arrives, it looks for data from the write cache and the read cache in turn. If no data is found, it will be read directly on the corresponding disk, and HBlock will select the most appropriate node according to the network distance and node load. When it has finished reading, it will put the data in the read cache and finally return the data to the client.

Considering the principle of time locality, the data that has just been accessed may be accessed in the future, and the data goes into the read cache. Users can adjust the size of the read cache or turn it off directly according to different scenarios.

Considering the principle of spatial locality, the adjacent data that has just been read may also be read, so it is up to the prereader to determine whether the data needs to be read from the disk in advance. Through the pre-reader, HBlock greatly improves the performance of the client small IO sequential read, reduces the response delay of the client read request, avoids frequent small IO operations on the disk, and gives better play to the disk performance.

These are the various caching mechanisms behind the high performance of HBlock, and they do work well: the main performance indicators of block storage are the bandwidth of large IO and the IOPS of small IO. Compared with open source Ceph, from the perspective of bandwidth, the write performance of HBlock is stably better than that of open source Ceph. From the perspective of IOPS, the write performance of HBlock can reach 3-5 times that of open source Ceph.

Live broadcast trailer

9 / 6 / 19 / 19 / 00 / 19 / 9 / 6 / 19 / 9 / 6 / 19 / 9 / 6 / 19 / 9 / 6 / 19 / 9 / 6 / 19 / 9 / 6 / 19 / 9 / 6 / 19 / 9 / 6

How to ensure that the service is highly available, the data is highly reliable, the service is super easy to use, and saves money without losing data. The above considerations of cost, performance and security are important factors for enterprises to choose distributed storage products.

Around these issues, Wu Zhimin, an expert in storage research and development of Tianyi Cloud International Business Department, and Wei Wei, director of storage product line of Tianyi Cloud International Business Department, visited the live broadcast room, bringing the sharing of "HBlock actual combat, deep excavation of original storage technology highlights".

In order to thank the users for their support and company

Tianyi Cloud is now open to community partners and developers.

Provide upper limit 1PB, 1024 nodes

Permanent free private cloud block storage service

(storage resource activation system version 3.4 only)

It's here

Expert "one-on-one" guidance

Flexible and diverse "soft" services

The event will close on 2023.10.31

Don't hesitate

Want to know more first-hand information about HBlock?

Lock Tianyi Cloud Video account

19:00 on September 6th

We won't part until we meet again!

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.