Overhead of Switching IO Schedulers Dynamically in Linux 2.6 Kernel

Dear All,

I am masters student in Computer Science at the University of Texas at El Paso, USA. I am working on my thesis breastled "coarse-grained adaptation of i-o schedulers in Linux". To explain to you short, for my thesis, I want to demostrate using some heuristics (that I will develop) the dynamic adaptation of i-o schedulers based on the characteristics of workloads and then quantifying the performance outcome. For this thesis, I also would like to evaluate the overhead of switching the i-o schedulers at run-time. This step is very crucial for my work and I have some questions regarding this. I will really appreciate it, if anybody can throw me some light on these questions. I am not a kernel programmer and I am a begginner with respect to Linux Kernel stuff. Hence I would appreciate everyone's help very much in this.

* I have read many articles about Jens' Modular I-O Scheduler work and related discussions in many forums but I havent got a clear answer for my questions.!!

My Questions --------------------

I installed the latest patch for dynamic switching of i-o schedulers on my experimental system. I know that switching takes place perfectly. But we can clearly see that it takes some time to switch from one scheduler to another. Now, I would like to know

1. What is the overhead (in time) of switching the schedulers dynamically?.

2. What is it dependent upon? i.e. is the overhead a function of no. of requests in dispatch queue i.e does it depend on no. of requests or anything else? or is it very neglibile ?

3. What exactly happens when I switch?. I mean what is the entry point and exit point in the code?. and what code is it?. i.e what happens exactly and where does "switching" actually take place in the code?. Is it in your patch itself?. At what place do you have the source code (the directory structure in the kernel).

4. Once I know the entry point and exit poiint, I can use a tool called kerninst (which is framework for dynamically splicing code into the kernel present at www.paradyn.org-html-kerninst.html ) to insert a timer function between the points. But Is there any other way,that you know of to time the overhead or the cost of switching?. I want to do this step to prove either the overhead is there or overhead is neglibile.

I really thank you all very much for your time.

Looking forward to your replies

Sincerely,

Zulu