{"id":97824,"date":"2020-10-22T14:42:26","date_gmt":"2020-10-22T12:42:26","guid":{"rendered":"https:\/\/prohoster.info\/blog\/administrirovanie\/huawei-oceanstor-dorado-18000-v6-v-chyom-eyo-haj-endovost"},"modified":"2020-11-18T00:58:50","modified_gmt":"2020-11-17T22:58:50","slug":"huawei-oceanstor-dorado-18000-v6-v-chyom-eyo-haj-endovost","status":"publish","type":"post","link":"https:\/\/prohoster.info\/en\/blog\/administrirovanie\/huawei-oceanstor-dorado-18000-v6-v-chyom-eyo-haj-endovost","title":{"rendered":"Huawei OceanStor Dorado 18000 V6: what makes it high-end","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

We thoroughly argue what makes the OceanStor Dorado 18000 V6 truly a high-end storage system with a solid foundation for the coming years. At the same time, we dispel common concerns regarding All-Flash storage and demonstrate how Huawei maximizes their potential: end-to-end NVMe, additional caching on SCM, and a whole array of other solutions.
\n\"Huawei
\n<\/a><\/noindex>
\n\"Huawei
\n<\/p>\n

A New Data Landscape \u2014 New Data Storage<\/h2>\n

\nData workloads are increasing across all industries, with the banking sector being a prime illustration. Over the past few years, the number of banking transactions has more than tenfold increased. As shown by BCG research<\/a><\/noindex>, in Russia alone, the number of cashless transactions using plastic cards rose more than thirty times from 2010 to 2018 \u2014 from 5.8 to 172 transactions per person per year. This is primarily due to the triumph of micropayments: most of us have become accustomed to online banking, and our bank is now at our fingertips \u2014 in our phones.<\/p>\n

The IT infrastructure of a credit organization must be prepared for such a challenge. And it is indeed a challenge. Among other things, whereas a bank used to require data availability only during business hours, now it must be 24\/7. Just recently, a latency of 5 ms was considered acceptable, but now even 1 ms is excessive. For a modern data storage system, the target value is 0.5 ms.<\/p>\n

The same applies to reliability: in the 2010s, there was an empirical understanding that reaching a reliability level of 'five nines' \u2014 99.999% \u2014 was sufficient. However, this understanding has become outdated. As of 2020, it is entirely normal for businesses to demand 99.9999% reliability for storage solutions and 99.99999% for the architectural solution as a whole. This is not a whim but an urgent necessity: there is either no maintenance window for the infrastructure or it is minuscule.<\/p>\n

\"Huawei
\n
\nTo visualize the figures better, it is convenient to project these indicators onto a monetary plane. The simplest way is to use an example from financial institutions. The diagram above shows how much each of the top 10 global banks earns in one hour. The Industrial and Commercial Bank of China alone makes a staggering $5 million during that time. This accounts solely for the loss of profit due to IT infrastructure downtime at the largest credit organization in China! From this perspective, it is clear that reducing downtime and enhancing reliability are not just justified by a few percentage points\u2014 even the smallest fractions are entirely rational. This is not only to improve competitiveness but also simply to maintain market positions.<\/p>\n

Comparable changes are occurring in other industries. For example, in air transportation: before the pandemic, airline services were picking up speed year after year, and many began to use them almost like taxis. As for consumer patterns, society has developed a habit of total service accessibility: upon arriving at the airport, we require Wi-Fi access, payment service connectivity, a map, etc. Consequently, the load on infrastructure and services in public spaces has increased manifold. The approaches to building this infrastructure that we considered acceptable even a year ago are rapidly becoming outdated.<\/p>\n

\"Huawei
\n<\/p>\n

Is it too early to switch to All-Flash?<\/h2>\n

\nTo address the tasks mentioned above, from a performance perspective, AFA\u2014All-Flash Arrays, which are completely built on flash storage\u2014are the most suitable. Until recently, doubts remained regarding whether they were as reliable as those based on HDDs and hybrids. After all, solid-state flash memory has a metric known as Mean Time Between Failures (MTBF). The degradation of cells due to input-output operations is, unfortunately, a reality.<\/p>\n

Thus, the prospects of All-Flash were overshadowed by the question of how to prevent data loss in case the SSD decides to give up the ghost. Backup is a familiar option, but the recovery time would be unacceptably long based on modern requirements. Another solution is to establish a second tier of storage on spindle drives; however, this approach loses some of the advantages of a strictly flash-based system.<\/p>\n

However, the numbers tell a different story: statistics from digital economy giants, including Google, over the past few years show that flash is significantly more reliable than hard drives. This applies to both short and long term: on average, it takes four to six years for flash drives to fail. In terms of data retention reliability, they are on par with or even surpass magnetic spindle disk drives.<\/p>\n

\"Huawei
\n
\nAnother traditional argument in favor of spindle drives is their cost-effectiveness. There's no denying that the cost of storing a terabyte on a hard drive is still relatively low. If we only consider equipment expenses, keeping a terabyte on a spindle drive is cheaper than on a solid-state drive. However, in financial planning, it\u2019s important to consider not only the purchase price of a specific device but also the total cost of ownership over time\u2014ranging from three to seven years.<\/p>\n

From this perspective, everything looks quite different. Even if we set aside deduplication and compression, which are typically applied to flash arrays and make their operation more economically advantageous, there are factors such as rack space occupied by the drives, heat generation, and energy consumption. In these areas, flash outperforms its predecessors. As a result, the TCO of flash-based storage systems, taking into account all parameters, is often nearly half that of arrays on spindle drives or hybrids.<\/p>\n

According to reports from ESG agency, All-Flash storage systems like the Dorado V6 can achieve a reduction in total cost of ownership (TCO) by up to 78% over a five-year period \u2014 largely due to efficient deduplication and compression, along with low energy consumption and heat output. The German analytics company DCIG also recommends these systems as the optimal choice for TCO among those available today.<\/i><\/p><\/blockquote>\n

The use of solid-state drives allows for space savings, reduces failure rates, decreases maintenance time, and lowers energy consumption and heat generation of storage systems. It turns out that AFA is at least economically comparable to traditional spindle-based arrays, and often surpasses them.<\/p>\n

\"Huawei
\n<\/p>\n

Royal Flash from Huawei<\/h2>\n

\nAmong our All-Flash storages, the top position is held by the hi-end system OceanStor Dorado 18000 V6. Moreover, it holds an industry speed record of up to 20 million IOPS in maximum configuration. Additionally, it is extremely reliable: even if two controllers fail at once, or up to seven controllers sequentially, or an entire engine goes down \u2014 the data will remain intact. Significant advantages of the 'eighteen-thousand' are provided by the built-in AI, including flexibility in managing internal processes. Let's look at how all this is achieved.<\/p>\n

\"Huawei
\n
\nA significant part of the advantage for Huawei lies in the fact that it is the only manufacturer in the market that builds storage systems entirely by itself. We have our own circuitry, firmware, and service maintenance.<\/p>\n

The controller in OceanStor Dorado systems is built on Huawei's proprietary Kunpeng 920 processor. It utilizes the Intelligent Baseboard Management Controller (iBMC) module, which is also our product. The AI chips, specifically Ascend 310, optimize failure predictions and provide tuning recommendations, and they too are Huawei's, as well as the input\/output boards \u2014 the Smart I\/O module. Finally, the controllers in solid-state drives are designed and manufactured by our team. All this provided the foundation for creating an integrally balanced and high-performance solution.<\/p>\n

\"Huawei
\n<\/p>\n

Over the past year, we have implemented a project for the deployment of our top-of-the-line storage system in one of the major Russian banks. As a result, more than 40 OceanStor Dorado 18000 V6 units in a metro cluster demonstrate stable performance: each system achieves over a million IOPS, even accounting for delays due to distance.<\/p><\/blockquote>\n

\"Huawei
\n<\/p>\n

End-to-End NVMe<\/h2>\n

\nThe latest Huawei storage systems support end-to-end NVMe, which is why we place significant emphasis on it. Traditionally used access protocols for drives were developed in the ancient days of IT: at their core are SCSI commands (hello, 1980s!), which bring along numerous functions to ensure backward compatibility. Whatever access method you take, the protocol overhead in such cases is colossal. As a result, storage systems that utilize SCSI-bound protocols have input-output latency that cannot be lower than 0.4\u20130.5 ms. In contrast, NVMe\u2014Non-Volatile Memory Express\u2014designed for operability with flash memory and free from the crutches of infamous backward compatibility, reduces latency down to 0.1 ms, not just on the storage system but across the entire stack, from host to drives. It's no surprise that NVMe aligns with the trends of data storage development for the foreseeable future. We have also bet on NVMe and are gradually moving away from SCSI. All modern Huawei storage systems, including the Dorado lineup, support NVMe (however, end-to-end implementation is only available on the advanced Dorado V6 series models).<\/p>\n

\"Huawei
\n<\/p>\n

FlashLink: a handful of technologies<\/h2>\n

\nA cornerstone technology for the entire OceanStor Dorado lineup is FlashLink. More accurately, this term encompasses an integral set of technologies aimed at ensuring high performance and reliability. These include deduplication and compression technologies, the operation of the RAID 2.0+ data distribution system, the separation of 'cold' and 'hot' data, and full-stripe sequential data writing (random writes, with new and modified data, are aggregated into a large stack and written sequentially, which enhances read-write speed).<\/p>\n

In addition to everything else, FlashLink includes two important components\u2014Wear Leveling and Global Garbage Collection. These deserve special attention.<\/p>\n

In fact, any solid-state drive is a miniaturized storage system with numerous blocks and a controller that ensures data availability. This is facilitated by transferring data from 'failed' cells to 'healthy' ones. This way, it is guaranteed that the data can be read. Various algorithms exist for this type of transfer. Generally, the controller aims to balance the wear across all cells of the drive. However, this approach has a downside. When data is moved within the SSD, the number of input-output operations drastically decreases. Currently, this is an unavoidable evil.<\/p>\n

Thus, if there are many solid-state drives in the system, the performance graph exhibits a 'sawtooth' pattern, with sharp peaks and drops. The problem is that any single drive in the pool can initiate data migration at any moment, while the overall performance is affected simultaneously across all SSDs in the array. However, engineers at Huawei have figured out how to prevent the occurrence of the 'sawtooth'.<\/p>\n

Fortunately, both the controllers in the drives and the storage controller, as well as the microcode at Huawei, are 'native'; these processes in the OceanStor Dorado 18000 V6 are launched centrally and synchronously across all drives in the array. Moreover, this occurs on the command of the storage controller and specifically when there is no significant input-output load.<\/p>\n

The artificial intelligence chip also participates in determining the right moment to transfer data: based on the statistics of accesses over the past few months, it can highly likely predict whether there will be an active input-output in the near future. If the answer is negative, and the system load at the moment is low, the controller commands all drives: those needing Wear Leveling should perform it all at once and synchronously.<\/p><\/blockquote>\n

Additionally, the system controller perceives what is happening in each cell of the drive, unlike the storage systems of competing manufacturers: they are forced to procure solid-state media from third-party vendors, which means cell-level detailing is unavailable to the controllers of such storages.<\/p>\n

As a result, the OceanStor Dorado 18000 V6 experiences a very short performance loss period during the Wear Leveling operation, which mainly occurs when no other processes are interfering. This provides consistently high and stable performance.<\/p>\n

\"Huawei
\n<\/p>\n

What Contributes to the Reliability of OceanStor Dorado 18000 V6<\/h2>\n

\nIn modern data storage systems, four levels of reliability are distinguished:<\/p>\n