By PURPLELEC | 26 October 2023 | 0 Comments
What architecture does industrial flash memory use to manage metadata?
The obvious features of industrial flash memory are stable performance, low latency and high random IOPS. For industrial flash memory, when evaluating performance, we generally focus on 90% of IO falling within the specified latency range (performance is a linear range, not a certain point). All software features such as data protection are implemented based on Inline, such as Inline deduplication, compression, and Thin-Provisioning.
1. Industrial flash memory architecture:
Scale out capability of flash memory: Horizontal scalability is the most basic feature to cope with concurrent access and improve performance capacity, so flash memory is a must-have function. Currently XtremIO supports 16 controllers, and solidfire already supports 100 controller nodes.
Controller symmetric A/A capabilities: The main application scenarios of flash memory are such as OLTP. Traditional A/P and ALUA arrays require time to switch when the main controller fails to switch over, causing IO to return to zero; and in flash memory arrays, generally The system CPU is the bottleneck, so only an A/A symmetric architecture with no ownership and balanced performance can better cope with it.
2. Metadata management
The design of industrial flash memory mainly considers how to bring out the random access performance of SSD. Unlike HDD, it is necessary to improve the order of the disk through prefetching and IO aggregation technology, and reduce the operation of mechanical disk to improve performance. Therefore, when designing industrial flash memory, we must consider how to optimize metadata (system metadata, deduplication and compression fingerprints, FTL mapping, etc.) management, IO scheduling strategies, garbage collection and wear leveling and other designs.
The two-layer metadata management architecture is a trend in realizing metadata management. The basic idea is that metadata mapping is based on the Map form of LBA->Block ID->Block location. The data LBA of the volume is mapped to the block ID instead of the physical location on the disk. Block address; after the data changes, you only need to change the corresponding block ID mapping relationship, and the block ID can be mapped to a new physical space. This simplifies the implementation and efficiency of deduplication and compression compared to the single-layer method.
SolidFire uses a two-layer metadata management architecture to implement metadata management. Metadata management adopts the Key-Value method; metadata mapping is based on the Map form of LBA->Block ID->Location. The data LBA of the volume corresponds to the block ID, not on the disk. The physical address; therefore, after the data changes, calculating the fingerprint only needs to change the corresponding block ID mapping relationship, which naturally supports deduplication. During garbage collection, the block marking method is used to clear unused blocks.
Leave a Reply
Your email address will not be published.Required fields are marked. *