Episode 5.16 - Optimization of Communication- Offload
Table of Contents:
00:27 - Optimization of Communication
00:44 - PCIe bandwidth (2nd gen) ~ 7GB/s
00:54 - 1000 FLOP/word to justify offload
01:23 - Large problems - good for offload
02:04 - Small problems - segnificant offload overhead
02:13 - O(n log n) problem offload
02:56 - Offload latency/bandwidth from data size
03:19 - Review: alloc_if/free_if
03:48 - Standard offload has penalties for allocation/deallocation
04:00 - Latency and Bandwidth with memory/data retention
04:27 - Overlapping Communication and Computation
04:44 - Review: asynchronous offload
05:03 - Standard approach to offloading tasks
05:21 - How Double Buffering works?
06:04 - Implementation of Double Buffering
06:17 - Double Buffering can be used for other problems
06:27 - Performance Results
06:40 - In the next episode: MPI