Fujitsu Develops In-Memory De-Dupe Technology

Fujitsu Laboratories Ltd. announced the development of a high-speed in-memory data deduplication technology for all-flash arrays, which are large-scale, high-speed storage systems and use multiple flash devices such as SSDs.

This technology enables the production of storage systems with up to twice the response speed when writing data, compared to previous methods.

In recent years, all-flash arrays have incorporated deduplication technology that consolidates duplicate data into one to write to a flash device, in order to utilize the limited capacity of flash devices. However, as the system connects to multiple flash devices through a network in order to search for duplicate data each time it writes data, and storage devices grow in capacity and increase in speed, a problem of lowered response speed during write operations arises.

Company’ labs has developed a method that can accelerate response speeds by executing deduplication after writing data. In addition, as data may be written to memory twice in some cases when processing is continued with the new method, thereby increasing communications volume and lowering overall processing performance, Fujitsu Labs has developed technology to automatically switch between the new method and the previous method, as operational conditions require.

This means that response speeds can be increased by up to two times, improving the response of virtual desktop services and reducing database processing times.

Details of this technology was announced at 28th Computer Systems Symposium (ComSys2016), held November 28 at Hosei University in Tokyo, Japan.

Development background
In recent years, with the advance of virtualization and the use of big data in business, performance requirements for storage systems have been increasing. All-flash arrays, which can increase performance by using high-speed flash devices and connecting multiple devices with a high-speed network, have therefore become widely used.

Flash devices are more expensive than traditional HDDs, and can only be written to a limited number of times over their lifetime, so all-flash arrays incorporate deduplication functionality that consolidates data to write only once to the flash device, which can increase the actual volume and reduce the number of unnecessary write operations

All-flash array storage system and deduplication (Figure 1)
Click to enlarge

Issues
In order to utilize expensive flash devices effectively, in-memory deduplication technology was developed for all-flash arrays featuring deduplication functionality. However, because previous technology would institute a process to search for duplicate data before writing through a network that connects multiple flash devices, with storage systems that grow in capacity and increase in speed it created issues with longer response times when writing data due to the overhead required for search processing.

About technology
The firm’ labs has developed a technology that can improve response speed for in-memory deduplication up to two times.

In-Memory deduplication technology (Figure 2)

New deduplication technology improves response speed under low load
With the new deduplication technology, the storage devices that make up the all-flash array cooperate to complete a write operation by temporarily storing the write data in open memory (the cache), improving response times. Then, while the server is preparing for the next write operation depending on the response from the memory, the technology searches for duplicate data, communicating between the storage devices in parallel, and the data is finally written to the SSD after duplicates are eliminated. With this technology, response times for write operations can improve by up to two times when the storage system’s network load is low.

Response time optimization technology actively switches
between two deduplication methods
Deduplicating after responding improves response speed compared to previous methods when the storage system’s network load is low, but when the load is high the network becomes congested. This reduces response times more than previous methods due to data writes across multiple devices taking longer (Figure 3). With the new response time optimization technology, of the two deduplication methods, the original method that writes after deduplication, for which the response time does not greatly vary, and the new method, for which response times vary widely, the system automatically chooses the method with the shortest response time by calculating the expected value for the average response time, based on records measuring the actual response time.

Network load leading to drop in write response speed (Figure 3)
Click to enlarge

With this technology, by automatically choosing the optimal method in response to the continually changing status of the system, it is possible to reduce system response time as a whole (Figure 4).

Response performance improvement
with newly developed method (Figure 4)
Click to enlarge

Effects
Using the newly developed technology, the company labs was able to achieve a lowest latency of about half that of previous technology in fio ⁽¹⁾ benchmarks. As a result, the lab was able to increase response speed when writing data to an all-flash array by up to two times For example, in applications such as virtual desktops and data base processing that require high write speeds, as accessing small files occurs in enormous volumes, there are many duplications. In such situations, user applications on the service could be sped up, improving the user experience. In addition, by applying this system to back-end storage for operations databases, operations systems could be sped up, enabling further consolidation of IT infrastructure.

Future plans
Fujitsu Laboratories will continue development of technologies to further accelerate all-flash arrays going forward, aiming to incorporate them into Fujitsu Limited‘s storage products from fiscal 2017 or later.

⁽¹⁾ fio: flexible I/O tester
A benchmark tool used to measure storage performance. It generates R/W load based on designated settings, and produces detailed performance data output.