PhD, École Polytechnique Fédérale de Lausanne
The Mondrian Data Engine: Unleashing the Vast Internal Bandwidth of Near-Memory Architectures for Data Analytics
The increasing demand for extracting value out of ever-growing data poses an ongoing challenge to system designers, a task only made trickier by the end of Dennard scaling. As the performance density of traditional CPU-centric architectures stagnates, advancing compute capabilities necessitates novel highly efficient architectures. Near-memory processing (NMP) architectures are reemerging as a promising approach to improve computing efficiency through tight coupling of logic and memory. NMP architectures are a great fit for data analytics, as they provide immense bandwidth to memory-resident data and dramatically reduce data movement, the main source of energy consumption. While moving from CPU-centric to NMP architectures alone boosts the performance of data analytics, maximizing NMP efficiency in terms of performance/watt is not equally straightforward. CPU-optimized data analytics operators rely on random memory access, which, in the context of NMP, result in wasteful DRAM row buffer activations. In addition, sustaining high enough memory-level parallelism to utilize NMP's ample bandwidth with fine-grained accesses requires non-trivial hardware, that cannot be accommodated under NMP's tight area and power constraints. We argue that efficient NMP calls for an algorithm-hardware co-design that favors algorithms with sequential accesses to enable simple hardware that accesses memory in streams. We then introduce the Mondrian Data Engine, an NMP architecture for data analytics that strikes that sweet spot.
Nooshin Mirzadeh is a student at Computer Sciences, École Polytechnique Fédérale de Lausanne (EPFL). She has been working in the Parallel Systems Architecture (PARSA) group, which advised by Prof. Babak Falsafi, since Sept. 2013. Her research interest lies in computer architecture, especially in high performance energy-efficient memory systems including 3D integration, and near-memory processing.