With data growing at an exponential rate, data centres face with complex challenges such as performance, scalability, high energy consumption, physical estate space and costs. To address the challenges faced by data centres, we proposed data centre future architecture based on next generation NVM.
Next generation non-volatile memory (NVM) technologies such as Spin-Torque Transfer Magnetic RAM (STT-MRAM) and Resistive random-access memory (RRAM) have very low latency access time (in terms of nanoseconds), provide data persistent against loss of power, and the prospect for higher capacity scaling than current DRAM technology. Next generation NVM can be integrated as main memory, storage or in a hybrid manner in data centres of the future and provides solution to address the growing performance gap between data storage and compute/memory speed, and the possibility of reduce energy usage in computer memories. By integrating next generation Non-Volatile Memory (NVM) technologies into the current storage system architecture, the new system is able to achieve better operating efficiency in terms of performance, cost, capacity and power consumption.
Our research programme in NVM takes an integrated approach with research spanning from the device to system and application layers.
NVM programme: Holistic approach spanning from device to system research
The NVM research in DCT primarily focus on the mechanisms and techniques at both the hardware (NVM controller) and software (OS and file system stacks, and programming model) layers for optimum integration of next generation NVM into enterprise storage and computing systems to achieve systems with low latency and high throughput performance, scalable, and energy and cost efficient.
NVM-based Storage System for Future Data Centre
The proliferation of IoT, smart nation applications and service digital economies generate huge data which need to store and analyse efficiently to improve society lifestyle and also businesses. Current data store and data analysis are based on traditional storage and computing hierarchy: storing data on hard disk or SSD and analysing data in volatile DRAM. Real time data analysis or low latency, high speed transaction is difficult to achieve for traditional system because data has to be ingested from the much slower disk or SSD to volatile memory. Emerging Non-volatile Memory (NVM) has DRAM-like performance and disk-like persistency. However legacy storage stacks incorporate complex components in modern operating systems which diminish the performance advantage of NVM.
In this project, our researchers address above architectural challenges by optimizing the file system stacks and memory management to effectively integrate NVM into storage system architecture. Our initial prototype system test results show MySQL Database performance running on our optimized NVM system stacks can outperform standard system by 9x:
NVM such as STT-MRAM with DRAM-like performance makes it possible to use as DRAM replacement with disk-like persistency feature. To keep data consistency in memory space is non-trivial because memory writes may be reordered by CPU and memory controller. We develop a novel data structure NV-Tree and persistent in-memory store solution that are able to provide consistency that is cache-optimized to reduce CPU cacheline fluhes. Our implementation of persistent in-memory key-value store system can improve performance by up to 4.8X.