Paper Code: 811.101
Title: Effective Local Scheduling of Parallel Jobs 
PCorReferee: r
PCorRefereeCode: 527.169
Originality:        [3]
Technical merit:    [4]
Readability:        [3]
Relevance:          [4]
Overall rating:     [3]
REJECT         [ ]
MILD   REJECT  [X]
MILD   ACCEPT  [ ]
ACCEPT         [ ]
-*--**---***-*--**---***-*--**---***-*--**---***-*--**---***-*--**--
Comments to authors
-*--**---***-*--**---***-*--**---***-*--**---***-*--**---***-*--**--
[[[direct]]]

Good paper.  My only concern is that the simulation is limited to
a single, synthetic testcase.  I would really like to see how the various
local and co-scheduling schemes work with a real workload.


-*--**---***-*--**---***-*--**---***-*--**---***-*--**---***-*--**--
Main Topic of the paper:
-*--**---***-*--**---***-*--**---***-*--**---***-*--**---***-*--**--

The authors argue that with a realistic uniprocessor scheduler and
a two phase fixed spin blocking policy, local scheduling can achieve
performance similar to co-scheduling, even for fine grain data parallel
programs.  They use simulation of a synthetic workload to drive 
the development of an effective blocking policy.  

-*--**---***-*--**---***-*--**---***-*--**---***-*--**---***-*--**--
The primary contribution of this paper is:
-*--**---***-*--**---***-*--**---***-*--**---***-*--**---***-*--**--
[X] Theoretical
[ ] Methodological
[ ] Application
-*--**---***-*--**---***-*--**---***-*--**---***-*--**---***-*--**--
Additional comments for the eyes of the Program Committee only:
-*--**---***-*--**---***-*--**---***-*--**---***-*--**---***-*--**--
[[[direct]]]

If the authors' claim (that local scheduling can be as efficient as
co-scheduling even for fine grain data parallel programs) is true, it 
should be possible to significantly increase the utilization of a parallel
computer.  The paper does an excellent job of arguing for the claim.
Where it is weak is that it relies on a single synthetic workload as
simulation input.  Simulating performance with real workloads is, I believe,
necessary to prove the claim.

Even though the claim was not proven to my satisfaction, I think 
the paper would be a good match for sigmetrics.

-*--**---***-*--**---***-*--**---***-*--**---***-*--**---***-*--**--