1 / 10

Software Thermal Management of DRAM Memory for Multicore Systems (SIGMETRICS 2008)

Software Thermal Management of DRAM Memory for Multicore Systems (SIGMETRICS 2008). Authors: Jiang Lin, Hongzhong Zheng , Zichun Zhu, Eugene Gorbatov , Howard David, and Zhao Zhang Presenter: Daniel Cole. Overview.

reid
Télécharger la présentation

Software Thermal Management of DRAM Memory for Multicore Systems (SIGMETRICS 2008)

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Software Thermal Management of DRAM Memory for Multicore Systems (SIGMETRICS 2008) Authors: Jiang Lin, HongzhongZheng, Zichun Zhu, Eugene Gorbatov, Howard David, and Zhao Zhang Presenter: Daniel Cole

  2. Overview • First study of thermal management for memory subsystems on real machines • Assumes very warm environment where emergencies often occur resulting in performance loss • Results (bandwidth throttling as baseline) • Core gating: up to 24.2% improvement, 13.3% on average • DVFS: up to 23.9% improvement, 15.3% on average • DVFS improved power by 15.5% and energy by 22.7% • Combining both may give addition performance improvement of 5.7% • DVFS previously not thought to have major impact

  3. A bit of background • In the latest generation of servers (2008), DRAM power can be comparable to processor • DRAM dynamic power is almost proportional to memory throughput

  4. Methods

  5. Methods • Core gating: • Lowers the temp by lowering cache misses, if the workloads are memory bandwidth-bound, the lower cache contention actually offsets the reduced computation • Thermal constraints already put a limit on memory bandwidth anyway • Note that active threads are migrated • DVFS • Reduce processor heat, which means less processor heat is being dissipated onto DIMMs • Note that this is uniform scaling on all cores

  6. Results

  7. Results

  8. Results

  9. Results

  10. Points of Note • The processor’s effect on memory temp is because the processor is usually very close to the memory DIMMs • In the machine where processor and DIMMs were closer DVFS helped more • DVFS doesn’t seem to help cache misses (makes sense) • Can’t seem to combine DVFS lower temp and gating cache miss improvement to get the benefits of both (only 5.7% on a certain machine) • Core gating doesn’t seem to reduce power at all (energy must be through faster execution: with no speed scaling, faster execution = lower energy)

More Related