Re: [ckrm-tech] Re: [patch 0/8] CKRM: Core patch set

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

Diego Calleja wrote:
> El Tue, 29 Mar 2005 22:05:30 -0800,
> Paul Jackson <pj@engr.sgi.com> escribió:
> 
> 
> 
>>worth having.  I for one am a CKRM skeptic, so won't be much help to you
>>in that quest.  Good luck.
>>
>>I don't see any performance numbers, either on small systems, or
>>scalability on large systems.  Certainly this patch does not fall under
>>the "obviously no performance impact" exclusion.
> 
> 
> I'm one of those people who also thinks that CKRM tries to do too much things, and
> although my opinion doesn't counts a lot, I'll try to explain myself anyway :)
> 
> One of the things I personally don't like about CKRM its how it handles "CPU resources".
> The goal of CKRM seems to be "control how much % a process can get get", but the
> amount of concepts created to achieve that is too huge and too complex. 


Certainly there's scope for improvement in the implementation of the CPU 
controller but the solution you propose works by redefining the problem.

> For the
> "CPU resources", I think that there're much simpler and better solutions. For example,
> instead what CRKM proposes I propose a simpler concept: "attaching" GIDs to a 
> niceness level.

Doing performance isolation at the granularity of users and groups may 
be useful but is not enough for workload management needs. There, it is 
essential that a a) flexible b) dynamic grouping of processes be 
controllable in their resource consumption as an aggregate. Tying that 
grouping to user/groups will not suffice.

CKRM's definition of class can be made equivalent to a user or group but 
not vice versa. Hence the more generic classes are being used, rather 
than reusing groups/users.

Also, our earlier prototype for the CPU controller had shown a 
0.14-0.63us overhead which remained constant with increasing number of 
processes. While we don't have measurements for later versions, the 
overhead figures are by no means unacceptably high if one values the 
additional generality of CKRM's class (over groups/users).



> 
> Say, we "attach" group foo to nice level -5. All users who belong to group foo will have
> permissions to renice themselves to nice -5. If instead of that, group foo has been
> attached at nice level 15, all processes from users who belong to foo will be run at 15,
> and they won't be able to renice themselves even to the default priority (0)
 >
> This should be very easy to implement, and what's more important, it'd probably have
> zero performance impact at runtime - CRKM touches hot paths in the scheduler
> I think, this would just touch a few non-critical places - because we'd just use a existing
> concept.


> Sure, this can't guarantee that a group will get reserved exactly 57% of  the CPU, but I
> think that such level of detail is unnecesary 

For desktop users, perhaps. For server workload management, this level 
of detail is necessary. As stated earlier, CKRM's design satisfies both.

> - instead we let the kernel uses the
> standard internal mechanisms to do the dirty job based in the distinction between
> standard nice levels. (And we could get that level of detail just by modifying the
> scheduler algorithm and adding a range of -50...0...50 nice levels ;)
> 
> For the CPU resources, we already have nice levels. The existing algorithms can already
> handle priorities with them. CKRM alternative seems to be to add a second scheduling
> algorithm which in super-hot paths like the ones from sched.c are, it will probably have a
> performance impact. In my very humble opinion, I think we should reuse existing UNIX
> concepts and combine them to achieve some of the goals CKRM tries to achieve in
> a much simpler (unixy ;) way.

Not that other Unix's design decisions should influence Linux but every 
other enterprise UNIX has some equivalent of CKRM's classes available. 
So the design is far from being non-unixy :-)


-- Shailabh