Available Technology

Collective Memory Transfers for Multi-Core Processors IB-2013-086

Berkeley Lab scientists John Shalf and Georgios Michelogiannakis have developed a new hardware component that can dramatically increase efficiency and reduce energy consumption in high performance supercomputers and chip multiprocessors. Using this new technology, the researchers demonstrated a nearly 40% reduction in execution time for reading and writing distributed data arrays, while reducing by 2.2 times the amount of energy required. In addition, simple programming constructs are presented to easily provide access to the CMS capabilities. The system is tailored to the wide variety of high-performance computing applications with algorithms that use many processors for vast data sets. Implemented with inexpensive dedicated hardware, the collective memory-scheduling (CMS) engine takes charge of transferring distributed data arrays between a collection of processors and the memory. The design tackles a well-known problem: DRAM underperforms when it receives requests for large amounts of data that are not in sequential “address order.” By controlling the collective transfer, the CMS engine accesses the memory in address order and distributes each datum to the appropriate processor according to the mapping of the distributed array to the processors. By restoring sequential order to requests, the CMS engine effectively increases the main memory bandwidth, which decreases execution time and reduces the energy required of such operations. Other approaches attempt to optimize memory processing by using logic redesigns to establish transaction queues and schedulers; reordering requests in-flight to the memory controller; or software strategies to load more data into each access request, thereby reducing the number of requests. The Berkeley Lab system outperforms all these approaches.
: - Maximizes memory throughput - Reduces application execution time by up to approximately 40% - 2.2x lower memory power consumption
Internal Laboratory Ref #: 
Patent Status: 
Patent pending. Available for licensing or collaborative research.
Share to Facebook Share to Twitter Share to Google Plus Share to Linkedin