At Interop Tokyo 2020, the Huawei next-generation OceanStor Dorado All-Flash Storage was awarded the Best of Show Award Grand Prize. Following rounds of strict reviews by top industry experts and scholars, this is the second time in four years that OceanStor Dorado has won the coveted prize.
"This is the first time we've seen an AI module used within a storage product" said one of the experts responsible for the review, "With this built-in AI module, OceanStor Dorado can intelligently analyze workloads in real time to optimize the cache prefetch algorithm and read hit ratio, ultimately boosting the system performance by 20%."
Huawei takes the lead in combining storage systems with AI, an innovation that was praised unanimously by the review board. This breakthrough is a result of the unremitting efforts of Huawei R&D team. To discuss the story behind the development of the first storage device with a built-in AI module, we spoke to Zhang Peng, chief architect of Huawei OceanStor Dorado All-Flash Storage.
High-end storage products are designed to provide ultimate reliability. In this principle hides many opportunities. Thanks to artificial intelligence (AI), Huawei is reinventing high-end storage.
"Reinvigorating high-end storage requires something strong, something innovative, and that's why Huawei has invested heavily in AI algorithms" said Zhang Peng. In recent years, Huawei has introduced an array of intelligent technologies into the OceanStor Dorado series. For example, an AI module is used based on a machine learning framework to proactively analyze and learn the I/O patterns of many application models and continuously improve the read cache hit ratio.
"The fully interconnected and shared architecture of OceanStor Dorado is ideal for running AI technologies such as machine learning. A product that can learn and analyze the global I/O patterns and support intelligent storage operations will improve the system performance and efficiency." Zhang Peng added.
Built-In AI Module Learns I/O Patterns and Optimizes the Prefetch Algorithm to Improve System Performance
Read cache is a common acceleration method for a storage system. Data is prefetched from disks and stored in a more quickly accessible location, generally in Random Access Memory (RAM). The CPU searches the cache first for the required data, and when found, the CPU sends it to the front-end interface module, which then sends the data to the user. The CPU searches for disks only when it cannot find the required data in the cache. The ideal and highest-performance situation is that all data read requests find data in the read cache. Data, however, is disordered and tasks are random. The user has no idea on which data should be fetched in advance and put it in the read cache.
To solve this problem, Huawei OceanStor Dorado innovatively uses an AI plug-in to improve prediction accuracy.
Figure 1 AI module + intelligent algorithm
Then how does the AI module help improve the forecast accuracy? A storage system receives many pieces of data, called I/Os. Each I/O is sent by different services and related not only spatiotemporally but also semantically. To make those relations easier to understand, let's look at some day-to-day examples. A time relation example is that nine o'clock follows eight o'clock; a space relation is how Russia is north of China; and a semantic relation is how "the world's largest bear" is a highly likely supplement to "Polar bear is." Similar patterns can be found from I/Os, and it is our job to find those patterns and improve the prefetch accuracy.
The AI module in OceanStor Dorado uses an integrated self-tuning deep learning algorithm, which can quickly analyze and deeply mine all of the I/O data of upper-layer services from the spatiotemporal and semantics perspectives. When an I/O arrives, the chip immediately identifies the data to be accessed and instructs the CPU to quickly obtain the data to the read cache. In addition, it continuously learns the existing data in the background to further improve accuracy. The chip then evaluates key performance indicators such as the prefetch hit rate, waste rate, and latency, and then makes adjustments to further improve accuracy.
According to Huawei tests, the read cache hit ratio of Huawei OceanStor Dorado climbs from 19% to 69% with the same bandwidth (16 Gbit/s FC) and under the same test model (random read, I/O block size within 64 KB).
Summary
Like a hard-working student, the built-in AI module of OceanStor Dorado uses every minute to continuously accelerate the storage system performance and improve user experience.
Looking back over the past two years, Zhang Peng smiled with relief "The new-gen OceanStor Dorado is innovative, it's very unique in the industry. I'm very proud."
AI is reinvigorating high-end storage. It is because of this innovation that the Interop review board wowed by the first AI module in a storage system. OceanStor Dorado has set a new benchmark for integrated intelligence for storage products.
