@ARTICLE{maheswaran:HCW:1999,
  AUTHOR = {Muthucumaru Maheswaran, Shoukat Ali, Howard Jay Siegel,
	Debra Hensgen, and Richard F. Freund},
  TITLE   = {{Dynamic Matching and Scheduling of a Class of Independent
	Tasks onto Heterogeneous Computing Systems}},
  CONFERENCE = {Heterogeneous Computing Workshop},
  MONTH = {April},
  YEAR = {1999},
  PAGES = {30--44}
}

% 
% Dynamic mapping (matching and scheduling) heuristics for a class of 
% independent tasks using heterogeneous distributed computing systems are
% studied. Two types of mapping heuristics are considered: on-line and
% batch mode heuristics. Three new heuristics, one for batch and two for
% on-line, are introduced as part of this research. Simulation studies are
% performed to compare these heuristics with some existing ones. In total,
% five on-line heuristics consider, to varying degrees and in different,
% task affinity for different machines and machine ready times. The batch
% heuristics consider these factors, as well as aging of tasks waiting to
% execute. The simulation results reveal that the choice of mapping
% heuristic depends on parameters such as: (a) the structure of the
% heterogeneity among tasks and machines (b) the optimization requirements
% and (c) the arrival rate of tasks
%


@inproceedings{hummel.loadbalance_hetero,
   author = "Susan Flynn Hummel and Jeanette Schmidt and R. N. Uma and Joel
Wein",
   title = "Load-Sharing in Heterogeneous Systems via Weighted Factoring",
   booktitle = "Proceedings of the 8th Annual ACM Symposium on Parallel
Algorithms
and Architectures",
   month = Jun,
   year = 1996,
   pages = {318--328}
}

% Paper Summary
%
% This paper describes an extension of Flynn Hummel, et. al.'s {\em factoring}
% method of work allocation that was presented in previous papers. The extension
% considers the impact of having heterogeneous workstations instead of the
% strictly homogeneous cases studied earlier. Experimental results are given
% showing that weighting work assignments generated by the original factoring
% algorithm by a value proportional to the expected speeds of the processors
% produced better finishing times than a number of other algorithms on an
% application running over a network of Sun workstations. Comparisons
% were run against static allocation, guided self scheduling (GSS), various work
% stealing, weighted work stealing, and factoring. The second part of the paper
% presents analytical results derived by assuming processor iteration execution
% times can be modeled by random variables. The results prove some bounds on
% how well weighted factoring in a heterogeneous system performs relative to a
% homogeneous system, and how well different weights perform relative to the set
% chosen by weighted factoring in a heterogeneous system.

@ARTICLE{hagerup:JPDC:1997,
  AUTHOR  = {Torben Hagerup},
  TITLE   = {{Allocating Independent Tasks to Parallel Processors:
              An Experimental Study}},
  JOURNAL = {Journal of Parallel and Distributed Computing},
  VOLUME  = {47},
  YEAR    = {1997}, 
  PAGES   = {185--197}
}

% 
%  Paper summary
% 
% This paper discusses the scheduling of batches of independent 
% tasks on a homogeneous set of processors. There is a fixed over
% head to pay for each batch. The trade-off is there for between
% small batches (high overhead, good load balance at the end of the run)
% and large batches (low overhead, possible imbalance at the end
% of the run). The assumption is that there is no possible work
% stealing or task duplication. Task execution times are assumed
% i.i.d. with known expected values and variances. The paper presents
% an interesting survey of different scheduling techniques ranging
% from ``timid ones'' (self-scheduling: batches of size 1) to
% ``bold'' ones. A common idea to move away from self-scheduling is
% to start with large batches and decreases the batch size as
% the computation makes progress. The hope is that the earlier large
% batches cause low overhead and that the late small ones still achieve
% good load balance. The different strategies surveyed in the paper are
% {\it static chunking}, {\it self-scheduling}, {\it fixed-size chunking},
% {\it guided self-scheduling}, {\it trapezoid self-scheduling}, 
% {\it factoring}, % {\it TAPER}. The author introduces a new strategy, 
% the {\it Bold} strategy,whose justification and insight are rather 
% cryptic (the strategy's algorithm is partially defined by what the 
% future behavior of that strategy will be !). The simulation results 
% at the end of the paper tend to prove that the Bold strategy does 
% indeed perform well. However, the performance metric
% chosen is {\it average wasted time} and one may wonder about the relevance
% of such a metrix (what about the makespan for instance ?). Overall, the 
% mostly interesting part of the paper is the survey of the different
% techniques and the corresponding bibliography.





@inproceedings{coljanni,
  author = {Michele Colajanni and Philip S. Yu and Daniel M. Dias},
  title = {Scheduling Algorithms for Distributed Web Servers},
  booktitle = {Proceedings of the 17th International Conference on
    Distributed Computing Systems},
  month = {May},
  year = {1997},
}

%
% summary of this paper coming to a .bib file near you soon...
%


@INPROCEEDINGS{webb,
  AUTHOR  = {Darren Webb and Andrew Wendelborn and Kevin Maciunas},
  TITLE   = {Process Networks as a High-Level Notation for Metacomputing},
  BOOKTITLE = {Proc. of the Int. Parallel Programming Symposium (IPPS'99), Workshop on Java for Distributed Computing},
  MONTH = {April},
  YEAR = {1999},
}

% Paper Summary
%
% Process networks are described as an abstraction for metacomputing and
% their use is demonstrated in a Geographical Information System.  The first
% section describes a process network as a hybrid between the Kahn model of
% process networks (each node is a sequential lightweight process) and
% Stevens' Java implementation for dataflow networks (each edge is a FIFO
% bounded queue that can read and write tokens).  Scheduling is done in a work
% queue manner such that processes are started when there is data available in
% their queue (so that scheduling is dynamic).  Additionally, processes within
% a process network can execute concurrently and mapped to distributed
% computers.  Also, since process networks are DAGs, there is never deadlock.
% The authors also point out that process networks can be hierarchical.
% Section 3 describes metacomputing and how distributed process networks can
% simplify the interface for users.  Section 4 describes their application of
% process networks in PAGIS (Process network Architecture for GIS) which is
% implemented in Java.  PAGIS provides an interface to a repository of
% geographic images and a set of utilities that can be applied to it.
% The PAGIS architecture consists of a client (UI and communication system),
% server, and worker.  One of the future goals of the authors is to integrate
% this architecture to DISCWorld, "a high level, service approach to
% metacomputing".