Computer Laboratory

Technical reports

Operating system support for simultaneous multithreaded processors

James R. Bulpin

February 2005, 130 pages

This technical report is based on a dissertation submitted September 2004 by the author for the degree of Doctor of Philosophy to the University of Cambridge, King’s College.

Abstract

Simultaneous multithreaded (SMT) processors are able to execute multiple application threads in parallel in order to improve the utilisation of the processor’s execution resources. The improved utilisation provides a higher processor-wide throughput at the expense of the performance of each individual thread.

Simultaneous multithreading has recently been incorporated into the Intel Pentium 4 processor family as “Hyper-Threading”. While there is already basic support for it in popular operating systems, that support does not take advantage of any knowledge about the characteristics of SMT, and therefore does not fully exploit the processor.

SMT presents a number of challenges to operating system designers. The threads’ dynamic sharing of processor resources means that there are complex performance interactions between threads. These interactions are often unknown, poorly understood, or hard to avoid. As a result such interactions tend to be ignored leading to a lower processor throughput.

In this dissertation I start by describing simultaneous multithreading and the hardware implementations of it. I discuss areas of operating system support that are either necessary or desirable.

I present a detailed study of a real SMT processor, the Intel Hyper-Threaded Pentium 4, and describe the performance interactions between threads. I analyse the results using information from the processor’s performance monitoring hardware.

Building on the understanding of the processor’s operation gained from the analysis, I present a design for an operating system process scheduler that takes into account the characteristics of the processor and the workloads in order to improve the system-wide throughput. I evaluate designs exploiting various levels of processor-specific knowledge.

I finish by discussing alternative ways to exploit SMT processors. These include the partitioning onto separate simultaneous threads of applications and hardware interrupt handling. I present preliminary experiments to evaluate the effectiveness of this technique.

Full text

PDF (1.3 MB)

BibTeX record

@TechReport{UCAM-CL-TR-619,
  author =	 {Bulpin, James R.},
  title = 	 {{Operating system support for simultaneous multithreaded
         	   processors}},
  year = 	 2005,
  month = 	 feb,
  url = 	 {http://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-619.pdf},
  institution =  {University of Cambridge, Computer Laboratory},
  number = 	 {UCAM-CL-TR-619}
}