From: Tarik Nesh-Nash (tarikn_at_microsoft.com)
Date: Mon Jan 26 2004 - 18:20:01 PST
This paper argues that kernel-based threads and user-level threads are not effective, and proposes a new interface of the kernel and the user package that improves the performance and flexibility.
The first section introduces the concept of thread and discusses its performance advantage, then presents the dilemma of user-level thread and kernel-level thread. The purpose of the paper is stated as to increase threads performance by benefiting from both methods.
The paper next argues that user-level threads are better than the kernel-level threads, since it offers safety, performance and flexibility, yet inadequate kernel support hinders this by enabling poor system integration.
The paper then proposes an alternative by building a new kernel interface and user-level thread system. The kernel provides the user thread with its own virtual multiprocessor. And the application has total knowledge over its scheduling state. A scheduling activation mechanism is implemented at the kernel level to interact with user threads. Potential deadlocks and performance hits due to the critical sections are mitigated with a recovery system.
A prototype is presented and compared its performance to existing systems.
I am finding a pattern about some of these papers; HYDRA, for example, just reduces the kernel to mechanisms and leaves the policies to be implemented at the user level, or Exokernel where the performance and the flexibility is separated between the OS and the application level. This paper argues how to improve the performance by implementing separate components in the OS and the application level. I understand that all these papers are more fine tuning the current technology than creating new radical paradigms.
This archive was generated by hypermail 2.1.6 : Mon Jan 26 2004 - 18:20:51 PST