- Anil Kurmus
- Nikolas Ioannou
- et al.
- 2017
- WOOT/USENIX Security 2017
Overview
In this activity we co-design, in collaboration with the IBM FlashSystem Development teams, flash memory controller functions to power the IBM enterprise storage systems, based on state-of-art 3D QLC flash memory.
Activities
All-flash arrays (AFAs) are ubiquitous IT elements in data centers, mainly because they offer superior performance and reliability compared to storage systems based on spinning disks. State-of-the-art AFAs are typically based on 3D-TLC (triple level cell) Flash technology and are commonly built based on off-the-self commercial SSDs. Flash memory vendors have recently introduced 3D-QLC (quad level cell) Flash technology to the market. However, this technology is currently inferior in reliability and performance compared to 3D-TLC. For that reason, many storage vendors may be reluctant to adopt 3D-QLC in their enterprise-grade AFAs, despite the significant benefits the new technology offers in terms of storage capacity and cost reduction.
IBM is one of the first AFA vendors that introduced general-purpose 3D QLC flash into enterprise storage systems, made possible by powerful flash memory controller innovations, that were co-developed at IBM Research. However, the continuous evolution in flash memory technology dictates new, more powerful and intelligent controllers to overcome the limitations of denser flash and leverage its capacity advantages. The Research team in Zurich is designing algorithms and flash management schemes to overcome the reliability and performance limitations of QLC flash and enable a next-gen Flash-based storage system with unparalleled capacity and competitive cost.
Our activities are focused on the following topics:
- Characterize and qualify many-layer 3D-QLC Flash from multiple memory vendors
- Design new algorithms to enhance raw flash endurance
- Design algorithms to boost performance, including schemes for LPT Paging and caching
- Innovations for cost reduction, including controller modeling for test cycle shortening.
- AI inside the FCM for garbage collection, wear forecasting and parameter adjustments
- Data-driven insights & development based on data from FlashSystem field usage
- Real-time ransomware detection at the drive level using machine learning and computational storage functions