Question

There are three thread models that are used in thread scheduling implementations usually done by OS Kernels. One of them is the hybrid (M:N) model in which some N application threads are mapped to M kernel threads so that they can use up to M processors. There are pros and cons to this model. One of the advantages is that the languages that are based on this model will introduce a language level scheduler implementation that is responsible for management and scheduling the application-level threads.

• I was wondering if anyone knows any effort or a work that already has done this so that a language or library could take advantage of it?
• Considering for instance the fact that Kernel 2.6.23+ uses an algorithm called CFS for scheduling, do you think that this hybrid model would at all be a wise approach to invest on?

Answer 1

First of all read this: http://www.kegel.com/c10k.html#1:1

Linux uses 1:1 threading model starting from kernel 2.6 (NPTL Native Posix threading library) and today almost all OSes move to this model:

• FreeBSD starting from 7.0 if I’m not mistaken.
• Solaris starting for some version also moved to 1:1 – I don’t remember which one.

Once Linux had M:N model (this was in 2.4 NGPT) but 1:1 is generally superior.

The biggest problem with M:N model is:

1. Hard to implement
2. When using blocking system calls you actually need to notify somehow kernel to block only one user space thread and not kernel one
3. In age of multi cores you want to have as much kernel threads as you can.

One of them is the hybrid (M:N) model in which some N threads of an application are mapped to some M available processors.

Small correction – N application threads mapped to M kernel threads so they can use up to M processors.