C462: Experiences from Integrating Algorithmic and Systemic Load Balancing Strategies

 Answers to Referees Report

The answers to specific referees comments are listed below.

o The use of the word systemic - I would advise that the authors to explain their
usage of the word at the beginning of the paper.
>> See Page 2, paragraph 1

o Page 4, paragraph 2, prepending "in the above mentioned" twice appears a clumsy
style. I would prefer something like "presented in the survey [1]".
>> See Page 4, paragraph 2

o Page 4, paragraph 2, you mention "a distributed architecture design" twice in
this paragraph. You should explain what you mean by this.
>> See Page 4, paragraph 2

o Page 4, paragraph 2, sentence starting "Hector is designed to…" add plural to
monitor.
>> See Page 4, paragraph 3

o Page 4, paragraph 3, perhaps remove "the" before MPICH. 
>> See Page 5, paragraph 1

o Page 5, paragraph 3, change "certain size amount of" to "certain sized amounts
of work"
>> See Page 5, paragraph 3

o Page 7, section 3.1, change "In Hector paradigm" to "In the Hector paradigm".
>> See Page 8, paragraph 1

o Page 7, section 3.1, change "don't" to "does not"
>> See Page 8, paragraph 1

o Page 9, section 3.2, add details about network technology connecting the cluster
together.
>> See Page 10, paragraph 1

o Page 13, paragraph 2, Question, why does the MA separate threads handling this
message queues?
>> We are not suggesting separate threads to handle the separate sockets (queues). A single thread will process these queues, giving priority to Fractile_Ask queue. However, this is a proposed solution which has not been implemented. We are only suggesting here that creating two queues, and giving preference to Fractile_Ask queue, should result in improved performance. We are not suggesting any implementation detail at this point.

o Page 14, section 3.2.2, add "a"  "does not result in a cost increase"
>> See Page 14, paragraph 2

o Page 16, section 4, Add discussion of how Hectiling would be affected my a
heterogeneous platform, such as a PC cluster connected via various bandwidth
interconnects. I would like to see a fuller discussion of the cost of task
migration and its affects on performance of Hectiling.
>> See Page 17, paragraph 5

o Page 16, section 4, please add some words on how the authors see the Hectiling
paradigm being generalized to support other applications which could benefit from
this integrated load balancing strategy.
>> See Page 17, paragraph 6