At FMS 2024, Kioxia had a proof-of-concept demonstration of their proposed a brand new RAID offload methodology for enterprise SSDs. The impetus for that is fairly clear: as SSDs get quicker in every technology, RAID arrays have a significant drawback of sustaining (and scaling up) efficiency. Even in instances the place the RAID operations are dealt with by a devoted RAID card, a easy write request in, say, a RAID 5 array would contain two reads and two writes to completely different drives. In instances the place there is no such thing as a {hardware} acceleration, the info from the reads must journey all the best way again to the CPU and foremost reminiscence for additional processing earlier than the writes will be achieved.
Kioxia has proposed the usage of the PCIe direct reminiscence entry characteristic together with the SSD controller’s controller reminiscence buffer (CMB) to keep away from the motion of knowledge as much as the CPU and again. The required parity computation is finished by an accelerator block resident throughout the SSD controller.
In Kioxia’s PoC implementation, the DMA engine can entry your complete host handle house (together with the peer SSD’s BAR-mapped CMB), permitting it to obtain and switch knowledge as required from neighboring SSDs on the bus. Kioxia famous that their offload PoC noticed near 50% discount in CPU utilization and upwards of 90% discount in system DRAM utilization in comparison with software program RAID achieved on the CPU. The proposed offload scheme may also deal with scrubbing operations with out taking on the host CPU cycles for the parity computation job.
Kioxia has already taken steps to contribute these options to the NVM Express working group. If accepted, the proposed offload scheme shall be a part of a regular that might turn out to be extensively accessible throughout a number of SSD distributors.
