Can DRAM do more than just store data?

Colloq: Speaker: 
Vivek Seshadri
Colloq: Speaker Institution: 
Carnegie Mellon University
Colloq: Date and Time: 
Fri, 2015-10-09 10:00
Colloq: Location: 
Building 5700, L204
Colloq: Host: 
Jeff Vetter
Colloq: Host Email: 
vetter@ornl.gov
Colloq: Abstract: 
In today's systems, DRAM is used only as a storage device. Off-chip DRAM interfaces allow the memory controller to read and write data. As a result, any operation must first read the required data from DRAM and store the results back into DRAM. In this line of work, we observe that this model is very inefficient for certain key primitives in modern systems. And we ask the question, "Can DRAM do more than just store data?" In response, we propose three techniques that exploit the DRAM architecture to significantly improve the efficiency of three important operations. First, we propose RowClone, a mechanism to perform bulk copy and initialization (specifically zeroing) operations completely within DRAM. RowClone improves the performance and energy efficiency of these operations by an order of magnitude. Second, we propose Gather-Scatter DRAM (GS-DRAM), a mechanism to improve the efficiency of non-unit strided access patterns. GS-DRAM achieves near ideal memory bandwidth and cache utilization for power-of-2 strided access patterns. Finally, we propose a new substrate that exploits existing DRAM operation to perform bulk bitwise operations completely within DRAM. Our mechanism enables an order-of-magnitude improvement in the throughput of bitwise operations. In this talk, I will provide a brief tutorial of DRAM operation. I will then describe these three mechanisms in detail.
Colloq: Speaker Bio: 
Vivek Seshadri is a Ph.D. student at the Computer Science Department at Carnegie Mellon University. He is advised by Prof. Todd Mowry and Prof. Onur Mutlu. His research interests are primarily in the field of computer systems, with specific focus on designing efficient memory systems.