Accounting of Time in Linux with Hyper-Threading Technology

Abstract della tesi in inglese To avoid stalling the execution of instructions during the transfer of operands to/from memory, modern processors use the so-called HyperThreading (HT) technology. With HT, the processing elements (ALU, FPU, etc.) are shared by two or more hardware threads. Hardware threads have their own set of registers, pre-fetch pipeline and branch predictors (when present). At the OS level, HT appear as independent cores. However, the execution time of a task over an HT is impacted by the usage of the HW resources made by the HT sharing the same processing elements. In short, a task executing for one second on an HT can progress more if the sibling HT is idle than if it is running. In this work, it is proposed a patch to the Linux kernel in which the accounting of time is aware of the busy/idle status of the other HT. The goal is to enable the scheduler to make fair scheduling decisions. ​