Shulou(Shulou.com)06/02 Report--

In recent years, super-convergence technology has become a popular infrastructure solution in modern data centers. With its advantages such as simple structure and easy to expand, super-fusion has a wide range of applications, from VDI, R & D testing and other scenarios to the production environment.

Take SmartX as an example, at present, many customers in finance, manufacturing, commercial chains and other industries have used SmartX super-converged products to run production business load, and some users are evaluating the future migration of core business to super-converged environment. However, compared with the physical machine and the all-flash disk array scheme, users still have concerns about the performance delay of the super-fusion system based on the virtualization platform, so that the core system has not been integrated into the super-fusion system.

On the other hand, Intel Aoton technology, especially Aoton persistent memory, is bringing more possibilities to the super-fusion system with its breakthrough features such as low latency and persistent storage. Based on the core business needs of users, through the full utilization and optimization of Aoton persistent memory and the upgrade of SMTX OS, SmartX has launched a high-performance, ultra-low latency super-convergence solution based on SMTX OS and Intel Aoten persistent memory.

Aoton is a new storage technology introduced by Intel, which allows memory cells to address independently without transistors, and supports a stacked architecture, also known as 3D-XPoint. Aoten technology supports different forms of encapsulation, which can be used either as memory or as storage. Aoteng persistent memory is encapsulated in the form of memory, which can be inserted in the memory slot like ordinary memory.

Aoten persistent memory has extremely high performance and less than 1us latency (1/1000 of NAND Flash latency), second only to memory latency. Its maximum reading and writing speed is more than 1000 times that of ordinary SATA SSD (NAND Flash), and its performance is stable and the jitter is small.

Based on the high performance, low latency and non-volatile characteristics of Intel Aoton persistent memory, SmartX hopes to create an all-flash super-fusion solution with Optane DC Persistent Memory (PMEM) as cache and NVMe NAND SSD as storage medium. This scheme will greatly improve the latency of business systems, so that SMTX OS products can be applied to the core businesses with the most stringent latency requirements. Using this scheme, the minimum system of SmartX super converged all-in-one machine with three nodes can reach 1.2 million IOPS, and the IO delay of virtual machine can be reduced from ms level to us level.

SMTX OS is a super-converged product of SmartX software. SMTX OS provides users with integrated infrastructure services such as virtual computing, virtual storage and virtual network in a software-defined way and distributed architecture. SMTX OS can run on x86 commercial server hardware, deploy as a distributed cluster, interconnect multiple nodes through high-speed Ethernet, and support scale-out (scale-out).

SMTX OS has the following features:

Simple: SMTX OS computing and storage integration, concise architecture, easy to manage.

Easy to scale: SMTX OS is deployed in a distributed architecture, easy to scale, and supports the addition of nodes online to expand capacity and performance.

High availability: SMTX OS supports VM high availability, asynchronous backup, and even dual active clusters, providing different levels of business continuity protection.

Good openness: SMTX OS is open to both hardware and computing virtualization platforms. Virtualization not only supports SmartX native ELF (KVM-based), but also supports VMware and Citrix virtualization platforms, and supports the mainstream server hardware in the market.

In the existing SMTX OS software stack, an IO request passes through the computational virtualization layer, the network layer, and the storage layer.

IO requests bring additional performance overhead as they pass through each layer of the system, which comes from context switching, system interruptions, memory copies, and so on. In order to give full play to the performance of PMEM, the solution optimizes these three levels to achieve end-to-end IO acceleration. The biggest challenge facing this scheme is how to maximize the performance of the storage system in the case of limited CPU resources.

The minimum size of the solution is still 3 server nodes, and multiple nodes are interconnected through 25GbE (or 100GbE) RDMA network to form a SMTX OS super-convergence cluster.

How Optane persistent memory cache acceleration works

The IO of the virtual machine first writes to the local Optane persistent memory cache, and writes the IO copy to the Optane persistent memory cache on the remote node through the high-speed RDMA network, and waits for all replicas to be written into the persistent memory before returning to ensure data consistency. With the characteristics of low latency and high performance of persistent memory, the virtual machine business will have an unprecedented performance experience.

If the data on persistent memory becomes "cold", the system will automatically brush the data to NVMe SSD for storage, ensuring that only a small number of persistent memory devices can continue to accelerate the business. If the "cold data" is accessed again, the data will be transferred back from the NVMe SSD to the persistent memory, ensuring that the frequently accessed "hot data" can reside in the persistent memory for optimal response speed. At the same time, NVMe SSD also ensures that the data transfer speed is fast enough.

Access persistent memory asynchronously to maximize bandwidth with limited CPU resources

Writing data from memory to persistent memory devices requires the operation of memory copy. Memory replication is a serial operation, which does not release CPU resources until the task is completed, which consumes a lot of CPU performance. Therefore, if data synchronization is written to persistent memory by default, CPU resources will be heavily consumed by memory replication, unable to handle other tasks, and the overall performance is relatively low.

If you need to further improve the bandwidth of persistent memory, you must increase the number of CPU core, through multiple CPU cores to access concurrently to achieve greater bandwidth performance. However, this approach is contrary to the concept of super-fusion architecture, which should compress the cost of the system as much as possible and set aside more computing resources for virtual machines, rather than unlimited occupation of CPU resources.

In order to improve the bandwidth performance of persistent memory under the limited CPU resources, the SmartX technical team creatively introduced the IO/AT DMA engine to implement the mechanism of asynchronously writing persistent memory, which increased the maximum write bandwidth of a single CPU core to 10GB/s.

SMTX OS supports persistent memory devices with almost no additional CPU resources and achieves a 2.5x performance improvement.

Storage engine optimization

Taking advantage of the persistent memory access feature, the SMTX ZBS storage engine is optimized to improve the problems encountered by most all-flash products on the market.

SMTX ZBS each node runs an IO processing logic locally, called LSM (Local Storage Management), which is responsible for the management of local cache and storage devices.

From the above architectural improvements and performance metrics, we can see that the super-fusion scheme based on SMTX OS and Intel Optane persistent memory has the following highlights:

1. The scheme realizes end-to-end optimization from Hypervisor to storage network and storage media.

two。 With only a small amount of computing resources, the minimum system can reach or even exceed the mid-to-high-end all-flash array index, while the application latency is much lower than that of the mid-to-high-end all-flash disk array.

3. Compared with mid-and high-end disk arrays, SmarX super-converged products have great advantages in total cost of ownership, system flexibility and easy maintainability.

The super-fusion scheme based on SMTX OS and Intel Optane persistent memory will bring value to users in the following scenarios:

1. Transaction database applications are mainly high concurrency, low latency IO scenarios, and provide: a. End-to-end ultra-low latency and high concurrency performance. b. Simple and flexible system architecture.

two。 Data center consolidation, and provide: a. Higher performance leads to higher virtual machine density. b. Further simplify IT and reduce total cost of ownership.

As the earliest domestic manufacturer to invest and focus on super-convergence, SmartX has a top domestic super-convergence and distributed block storage research and development team, insists on independent research and development of core technologies rather than based on open source from the initial stage, and has a strong competitive advantage of differentiation. At the same time, SmartX always insists on local R & D and localization services, providing users with a series of original factory services, including installation, training, after-sales door-to-door, etc., which can provide users with efficient and reliable factory support to the maximum extent, so that users'IT infrastructure can be guaranteed by "360-degree" services.

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.