From: Shailabh Nagar (nagar_at_watson.ibm.com)
Date: Fri 07 Nov 2003 - 21:57:34 GMT
Rik van Riel wrote:
>On Fri, 7 Nov 2003, Sam Vilain wrote:
>>Have you got implementations for all these parts too?
>Except for the disk quota limits, it's all in CKRM.
>The CKRM people are also working on a fair IO scheduler
>and I plan to add some other things, like limiting the
>number of processes, file descriptors, etc... per class.
A first-cut of the CKRM I/O scheduler is available and on the
It only allows equal shares to all classes (just like the equal shares
to all processes provided by
the CFQ iosched).
A variable-bandwidth providing version which respects share settings as
best as it can is in the works.
Some other work put me off completing this but its conceptually quite
simple - instead of picking off requests in
round-robin off the per-class queues, the iosched picks them off using a
virtual clock based algorithm. This will provide
proportional serving of requests as far as the iosched is concerned (it
might not result in the same share of disk bandwidth being
allocated since that depends on the seek times seen by the requests but
it should be good enough to start with - modelling the disk etc.
to get more accurate apportioning will most likely add more complexity
than required and/or violate a few layering assumptions).
Vserver mailing list