Datacenter designers today are faced with a unique set of problems. With increasing cloud and machine learning applications—in addition to the end of Moore’s law—they face the challenge of designing systems that can support the intense strain on data requirements.
Worldwide data trend. Screenshot used courtesy of Kogod School of Business at American University
Databases, their infrastructure, and their applications have not been able to scale at sufficient rates to keep up with data demands. Pliops, a flash-based storage company, has been working to address these problems by exploiting inefficiencies in database applications and storage (SSDs).
Inefficiencies in Key-Value-Based Storage Engines
In a Pliops white paper, the company explains that key-value (KV)-based storage engines are commonly used to manage data persistence and indexing tasks common to transactional databases, analytics applications, and software-defined storage (SDS). Unfortunately, these storage engines introduce specific inefficiencies in the way they utilize SSDs, preventing scalability as a result.
Major inefficiencies in KV store operations are caused by large levels or read, write, and space amplification. Each can be defined as follows:
Example of write amplification. Image (modified) used courtesy of Music Sorter [CC BY-SA 3.0]
- Read amplification: Regardless of the requested read size, databases retrieve data in complete blocks from storage systems. In this way, small read requests can become much larger ones. Many flash-based applications amplify reads by 100 times.
- Write amplification: Storage engines either perform garbage collection or must write complete blocks when the data to be updated is much smaller than a block. This causes blocks of data to be written, erased, and rewritten multiple times.
- Space amplification: Databases can store variable-length data, but SSDs store data at fixed-lengths, which are often much larger than the natural storage unit of most applications. The methods used by the most common applications to store and index data require data structures to not be full, resulting in wasted storage space.
Pliops’ New Storage Processor
This week Pliops, a flash-based storage company, announced the release of its newest storage processor, which they claim solves these issues.
New storage processor. Image from Pliops
Pliops’ storage processor (PSP) is a hardware-based storage accelerator that enables cloud and enterprise customers to offload and accelerate data-intensive workloads.
Recently tested by more than 10 tier-one cloud and enterprise companies, the PSP was proven to increase performance by 10 times, reduce latency up to 1,000 times, and increase flash price/performance by more than 90%. These results maintained consistent amongst almost all of the flash workloads.
What Makes the PSP Work So Well?
The PSP benefits from a novel data structure enabled by hardware acceleration that allows for increased storage performance. Pliops claims the new dynamic storage structure allows the PSP to achieve high-performance levels by keeping amplification close to the optimal value.
Existing storage stack (left) and storage stack with the PSP (right). Image from Pliops
In a white paper, the company explains, “In general, Pliops Storage Processors works by eliminating the inherent inefficiencies present in databases and applications using those databases. When using software-based storage engines such as RocksDB, these inefficiencies cause significant read, write, and space amplification that consumes resources and slows performance.”
PCle Card Deployment
Coming in a PCIe card, the PSP offers an easy-to-integrate solution that doesn’t require changes to the underlying software. With strong test results proven by outside cloud companies, the PSP looks to be a promising solution to speed the scalability in data centers as demands continue to grow.
- Decrypted: Hackers show off their exploits as Black Hat goes virtual - August 15, 2020
- Arduino-controlled robot arm is ready to play you in chess - August 15, 2020
- Elon Musk says ’embarrassingly late’ two-factor is coming to Tesla app - August 15, 2020