Energy-Aware Interface for Memory Allocation in Linux

Student thesis: Master thesis (including HD thesis)

  • Martin Jensen
  • Lars Rechter
4. term, Software, Master (Master Programme)
In this master thesis, we extend the Linux kernel to support grouping frequently accessed (hot) and infrequently accessed (cold) data on different memory hardware. By doing this, the memory hardware with cold data can reduce energy consumption by going into low power states. We manage this separation in the kernel by adding an additional zone for cold data, which is adjustable at compile time. Processes can allocate memory in the cold zone with an extension to the mmap system call. We test the memory layout of our machine with the benchmark STREAM, showing that the modified kernel behaves as desired in terms of memory separation. Additionally, we implement a proof of concept in-memory database to benchmark the power consumption and run time performance of our modified kernel. The results show a smaller overhead than expected, but no reduction in power usage. We attribute the unchanged power usage to the memory power management strategy of the memory controller in our test machine.
Publication date28 Jun 2021
Number of pages78
ID: 414480356