Improving the user experience while reducing response times

Editorial Type: Opinion Date: 2021-02-01 Views: 158 Tags: Storage, NAS, SSD, Data Management, Data protection, Infortrend PDF Version:
Frank Lee, Senior Director of Product Planning at Infortrend, explains how to get the best out of a scale-out NAS solution using SSD caching and RDMA

Think about how you expect data to appear on your screen when making a read request: almost instantaneously. Nowadays, data access speed is increasingly important for all applications involving human interactions with massive amounts of data, whether it is media editing, massive backup of corporate files, healthcare PACS, or parallel high-performance computing. So, it is essential for better user experience and time-sensitive applications that storage provides rapid file access performance regardless of how much data has to be processed.

With this in mind, Infortrend developed a scale-out NAS product family, EonStor CS, that allows enterprises to experience faster response time, higher performance gains, and lowered I/O latency of CS cluster while processing the sheer volume of data. EonStor CS was specially designed for data-intensive networks of growing enterprises which deploy NAS cluster for storing massive amounts of unstructured corporate data.

WHY SCALE-OUT NAS CLUSTER - AND WHAT ARE THE CHALLENGES?
The amount of generated data is growing rapidly, and 80-90% of it comprises unstructured data such as emails, presentations, web pages, audio, and video files. The importance of unstructured data for enterprise success is difficult to overestimate, as it brings valuable insights and unleashes the whole new business perspectives. To satisfy these emerging requirements, organisations should ensure computing performance and capacity of the whole IT infrastructure. A scalable infrastructure allowing the addition of more high-performance storage systems and expansion enclosures turns out to be extremely important for growing companies to handle increasing workloads.

EonStor CS is a scale-out NAS solution that offers not only horizontal (scale-out), but also vertical (scale-up) expandability. By adding more CS nodes and expansion enclosures, enterprises can get a high performance and high-capacity cluster system with up to 100+ GBps Read/Write speed and more than 100PB of storage for the needs of future data growth. Scale-out expansion provides an easier and more cost-effective way of managing growing data while reducing consequent performance bottlenecks. Initial deployment can start with only one CS storage. Further, as an organisation's demands grow, more CS nodes can be added to linearly increase the system performance and capacity. CS also simplifies data access and management by integrating data from all nodes into one cluster system under a single namespace. In this way, the problem of data silos is eliminated, and all the data can be visible and easily accessible for users.

Being a perfect solution for storing large volumes of unstructured data, scale-out NAS, however, encounters some issues to be handled. In large networks, massive amounts of files often negatively impact node response times. And since there are two kinds of workloads (large streaming reads/writes and small random reads/writes), it is necessary for smooth enterprise workflow to provide a storage solution that can handle both well. Another factor that can also prolong the response time for users' I/O request is node communication latency appearing in clusters with multiple deployed nodes. To tackle these challenges, Infortrend implemented SSD cache and RDMA technology for its EonStor CS scale-out NAS.

WHY IMPLEMENT SSD CACHE IN SCALE-OUT NAS?
Enterprises can add up to four SSDs to each node of EonStor CS to cache file metadata as well as small-size files which are usually transferred in small-size random I/O. In other words, CS will store frequently used random access data in SSD cache to improve performance. Mechanical HDDs are not suitable for such a scenario since they have higher average seek time which causes high latency in random data access.

With SSDs' near-zero seek time advantage, SSD cache can improve random data access performance dramatically. The EonStor CS SSD cache also stores metadata. In the cluster storage architecture, for all the operations that clients may do to the file, such as list, open, close, rename, delete, read, and write, metadata needs to be accessed. Thus, the performance of accessing metadata determines the performance for file operations. Besides, the more nodes, directories and files in a cluster, the more metadata access will be required. To solve this problem, CS will tag the metadata and store it into SSD cache for speeding up file operations.

However, SSD also has its weaknesses. Random data write I/O workloads cause write amplification issues in SSD and, therefore, shorten its lifespan. To avoid this, CS collects the random write data in the RAM and writes it to the SSD cache sequentially afterwards. And for large sequential file I/O workloads, HDDs are a better fit as constantly storing sequential data on SSD cache will cause cache invalidation and shorten the SSD's lifespan. Besides, HDDs perfectly meet the performance requirements of large sequential workloads.

EonStor CS will be especially appealing for modern lean enterprises, as it is a considerably more cost-effective solution (compared to all-flash array) of high-performance storage with lower cost and high capacity for massive data volumes.

WHY RDMA BETWEEN SCALE-OUT NAS CLUSTER NODES?
As a cluster system, CS nodes use internal networking for nodes intercommunication. We implemented RDMA for internal networks of EonStor CS scale-out NAS to boost the overall cluster performance by up to 30%. For preventing data loss and ensuring data protection, CS is designed with support of node protection Erasure Code and Replica modes which are also executed between nodes occupying the bandwidth of the cluster internal network. With RDMA support, CS nodes' CPU are offloaded from data copies, and this allows cluster nodes to perform data I/O directly through the memory ('zero-copy networking' feature).

With lowered CPU load, significantly shortened I/O latency, and higher bandwidth, CS RDMA cluster system boosts network and host performance by leveraging its CPU for processing massive I/O workloads and increases IOPS.

Totally new operating experience By integrating SSD cache into CS nodes and implementing RDMA for node-to-node internal communication, Infortrend provides an outstanding scale-out enterprise NAS solution that is able to achieve faster response time, reduced latency, and, as a result, totally new and improved user experience. CS is a well-received solution by the Oil and Gas industry for instance, where a large-scale cluster is deployed by one customer (a big oil exploration and production company in China) for HPC processes in seismic analysis as well as for daily files operations.

Based on feedback from our customers in the Media and Entertainment industry with CS cluster deployed, media ingestion and processing of 4K and 8K ultra-high resolution video is fast and easy. "We have hundreds of thousands of media materials stored on CS systems, and we can feel that the time spent on browsing is greatly reduced," says the Discovery Channel. "When using Premiere Pro to playback multiple HD videos at the same time, the data access latency is shortened, and the problem of frame dropping when multiple users are playing back videos is also solved - and we like the price!"

More info: www.infortrend.com