[Devel] Re: [ckrm-tech] [RFC][PATCH 2/7] UBC: core (structures, API)

[Chandra Seetharaman]

On Fri, 2006-08-18 at 19:38 -0700, Matt Helsley wrote:

<snip>

> > > 
> > >>+	for (p = ub; p != NULL; p = p->parent) {
> > > 
> > > 
> > > Seems rather expensive to walk up the tree for every charge. Especially
> > > if the administrator wants a fine degree of resource control and makes a
> > > tall tree. This would be a problem especially when it comes to resources
> > > that require frequent and fast allocation.
> > in heirarchical accounting you always have to update all the nodes :/
> > with flat UBC this doesn't introduce significant overhead.
> 
> Except that you eventually have to lock ub0. Seems that the cache line
> for that spinlock could bounce quite a bit in such a hot path.
> 
> Chandra, doesn't Resource Groups avoid walking more than 1 level up the
> hierarchy in the "charge" paths?

Yes, charging happens at one level only (except the case where the group
is over its guarantee, it has to borrow from its parent, it will go up).

<snip>
-- 

----------------------------------------------------------------------
    Chandra Seetharaman               | Be careful what you choose....
              - sekharan at us.ibm.com   |      .......you may get it.
----------------------------------------------------------------------