Two years in the past we reported on IBM’s FlashCore Module, their customized U.2 NVMe SSD to be used of their FlashSystem enterprise storage home equipment. Earlier this 12 months IBM launched the FlashCore Module 2 and this week they detailed it in a keynote presentation at Flash Memory Summit. Like its predecessor, the FCM 2 is a really high-end enterprise SSD with some uncommon and shocking design selections.
The most uncommon function of the primary IBM FlashCore Module was the truth that it didn’t use any supercapacitors for energy loss safety, nor did the host system embrace battery backup. Instead, IBM included Everspin’s magnetoresistive RAM (MRAM) to supply an inherently non-volatile write cache. The FCM 2 continues to make use of MRAM, now upgraded from Everspin’s 256Mbit ST-DDR3 to their 1Gbit ST-DDR4 reminiscence. The higher-density MRAM makes it a lot simpler to incorporate a helpful amount on the drive, nevertheless it’s nonetheless far too costly to completely substitute DRAM on the SSD: managing the FCM2’s multi-TB capacities require a number of GB of RAM. IBM’s primary motivation for utilizing MRAM as a write buffer as a substitute of DRAM with energy loss safety is that supercaps or batteries are inclined to have service lifespans of just a few years, and when an vitality storage system fails issues can get ugly. IBM sees MRAM as providing higher long-term reliability that’s value the fee and complexity of constructing a drive with three sorts of reminiscence.
The FCM 1 used Micron 64-layer 3D TLC NAND, which on the time was a fairly customary alternative for high-end enterprise SSDs. The FCM 2 makes the daring swap to utilizing Micron’s 96L 3D QLC NAND. The increased density and decrease value per bit has enabled them to double the utmost drive capability as much as 38.four TB, however sustaining efficiency whereas utilizing inherently slower flash is a tall order. Fundamentally, the brand new NAND has about 3 times this system (write) latency and 2-Three occasions the learn latency. Write endurance and knowledge retention are additionally decrease. But the FCM 2 remains to be rated for two DWPD and IBM claims elevated efficiency due to a mix of a number of tips.
IBM’s FlashCore Modules use a customized SSD controller structure carried out on an enormous FPGA. The 20-channel NAND interface explains the marginally odd drive capacities in comparison with extra run of the mill SSDs with eight or 16 channel controllers. IBM contains line-rate clear compression derived from the {hardware} compression supplied on IBM Z mainframes. This gives a compression ratio round 2.3x on typical knowledge units, which fits a good distance towards mitigating the endurance points with QLC (however the FCM 1 additionally had compression, so this is not an enormous benefit for the FCM 2). The FCM 2 can also use a few of its QLC NAND as SLC. This is not so simple as the SLC write caches discovered on just about all client SSDs. Instead, the FCM 2 tracks IO patterns to foretell which chunks of knowledge will probably be often accessed (“hot” knowledge), and tries to retailer these on SLC as a substitute of QLC whereas sending “cold” knowledge straight to QLC. Enterprise SSDs sometimes keep away from utilizing SLC caching as a result of it makes it laborious to make sure good QoS throughout sustained workloads. (Client drives can rely on real-world workloads providing loads of idle time that can be utilized for cache flushing.) IBM appears assured that their sensible knowledge placement heuristics can keep away from any critical QoS points, and the FCM 2 drive may also make use of knowledge lifetime hints supplied by host software program.
Using the FCM 2 drives, IBB’s FlashSystem storage home equipment can provide 40GB/s per 2U/24 drive system, with usable capacities of as much as 757 TB or an efficient capability of about 1.73 PB due to the built-in compression.
