Facebook uses PCIe Flash to add power to cheap disk arrays

Http://www.aliyun.com/zixun/aggregation/1560.html ">facebook upgraded the Flashcache Open Source tool, allowing administrators to get higher performance from inexpensive disk arrays configured with PCIe flash cards.

The
Flashcache tool is now upgraded to version 3.0. The tool allows Facebook to use the high-performance cache on the PCIe Flash memory card to speed access to critical data without costly use of full flash arrays.


Flashcache is a write-back block caching technique that is deployed as a target for a Linux kernel device mapping, making it easy to use a general-purpose system to respond to high traffic applications.


Enterprise-Class flash memory plus a large array may be for those who want to build multiple TB, it is appealing to people who need a network access latency system, which does not need to rewrite the software to play its part, so the investment in a handful of devices is far less than the investment that puts everything on full flash, Mituzas, a Facebook data engineer, said.


Flashcache 3.0 provides a better read/write distribution, adjusts the size of disk and flash memory, and distributes thermal data to flash memory to avoid bottlenecks. Facebook has also modified its flash recovery and write efficiency technologies to provide more predictable performance.


, although originally designed for Facebook's own use, has been widely favored by the community. "We have seen the efforts of the community around Flashcache-mailing lists, open source code submissions, and consulting companies in the database field are beginning to provide support for Flashcache," says Domas Mituzas, a Facebook data engineer. ”

The next area of
Flashcache technology development includes metadata reconstruction to more efficiently speed up data access, ensuring that no excessive writes are made to the cache, preventing the underlying infrastructure from being overwhelmed by queued write operations.


Mituzas explains: "Finally, when we have terabytes of cache, dozens of TB of data per device, we need to balance memory and CPU usage carefully." More efficient CPU algorithms will require higher costs. For example, add more pointers or time markers for metadata entries, and if you use 2TB caching, a system page requires 4GB of RAM ... Because the application can make good use of it. ”


It's worth noting, however, that Facebook's tools are not for everyone, because you need to have some expertise and scale before you realize a fully integrated build stack.


Mituzas said: "From expensive technology to cheap technology, this requires a lot of software-related work--which can save a lot of costs on a large scale, and, on the other, require more powerful storage devices, so that you can move faster in storage-led system design. ”