A Migratory Near Memory Processing Architecture Applied to Big Data Problems
share
Servers produced by mainstream vendors are inefficient in processing Big Data queries due to bottlenecks inherent in the fundamental architecture of these systems. Current server blades contain multicore processors connected to DRAM memory and disks by an interconnection chipset. The multicore processor chips perform all the computations while the DRAM and disks store the data but have no processing capability. To perform a database query, data must be moved back and forth between DRAM and a small cache as well as between DRAM and disks. For Big Data applications this data movement in onerous. Migratory Near Memory Servers address this bottleneck by placing large numbers of lightweight processors directly into the memory system. These processors operate directly on the relations, vertices and edges of Big Data applications in place without having to shuttle large quantities of data back and forth between DRAM, cache and heavyweight multicore processors. This paper addresses the application of such an architecture to relational database SELECT and JOIN queries. Preliminary results indicate end-to-end orders of magnitude speedup.
READ FULL TEXT
