[Devel] Re: [ckrm-tech] [PATCH 4/7] UBC: syscalls (user interface)
pj at sgi.com
Fri Aug 18 11:56:24 PDT 2006
> In order to minimize this effect, resource controllers should be
> providing both minimum and maximum amount of resources available for a
> resource group.
No - not "should be." Rather "could also be."
The fair sharing model (such as in CKRM) that strives for maximum
utilization of resources respecting priorities and min/max limits is
(I suppose) quite useful for certain workloads and customers.
The hardwall NUMA placement model (such as in cpusets) that strives
for maximum processor and memory isolation between jobs, preferring
to leave allocated resources unused rather than trying to share them,
is also quite useful for some. Customers with 256 thread, one or
two day long run time, -very- tightly coupled huge OpenMP Fortran
jobs that need to complete within a few percent of the same time,
every runtime, demand it.
Don't presume that fair sharing -should- always be preferred to
hardwall NUMA placement.
Just not so.
Besides -- what benefit would CKRM gain from Andrew's latest
brainstorm? Doesn't CKRM already have whatever means it needs to
define and share pools of memory?
I won't rest till it's the best ...
Programmer, Linux Scalability
Paul Jackson <pj at sgi.com> 1.925.600.0401
More information about the Devel