Re: [RFC][PATCH] O(1) Entitlement Based Scheduler

[Shailabh Nagar <nagar@watson.ibm.com>]

Mike Fedyk wrote:

> Shailabh Nagar wrote:
>
>>>> Mike Fedyk wrote:
>>>>
>>>>> Better would be to have the kernel tell the daemon whenever a 
>>>>> process in exec-ed, and you have simplicity in the kernel, and 
>>>>> policy in user space.
>>>>
>>
>>
>>
>> As it turns out, one can still use a fairly simple in-kernel module 
>> which provides a *mechanism* for effectively changing a process' 
>> entitlement while retaining the policy component in userland.
>
>
> How much code could be removed if CKRM triggered a userspace process 
> to perform the operations required?


In CKRM, the code to perform classification is an optional  kernel 
module. So size isn't really an issue in terms of impact to core kernel 
code.

Our prototype version of the classification engine, RBCE, is about 2700 
lines without any effort being put into reducing its size etc. If that 
were to be completely pared down to only provide events to userspace, it 
would come down by quite a bit (can't say exactly how much but atleast 
50% is a safe bet).

I think the more important question is performance impact - what do you 
give up in terms of efficiency and granularity of control by going to 
userspace vs what you gain in reduced kernel pathlength. Empirically, we 
found RBCE was quite efficient  but no quantitative analysis was done.

-- Shailabh