5 <img align=center src="simgrid_logo.png" alt="SimGrid"><br>
9 \section overview Overview
11 SimGrid is a toolkit that provides core functionalities for the simulation
12 of distributed applications in heterogeneous distributed environments.
13 The specific goal of the project is to facilitate research in the area of
14 distributed and parallel application scheduling on distributed computing
15 platforms ranging from simple network of workstations to Computational
18 \section people People
20 The authors of SimGrid are:
22 \author Henri Casanova <casanova@cs.ucsd.edu>
23 \author Arnaud Legrand <arnaud.legrand@imag.fr>
24 \author Martin Quinson <martin.quinson@tuxfamily.org>
26 \section intro Available Softwares
28 The SimGrid toolkit is composed of different modules :
30 \li XBT (eXtensive Bundle of Tools) is a portable library with many
31 convenient portable datastructures (vectors, hashtables, heap,
32 contexts ...). Most other SimGrid modules rely on it.
34 \li SURF provides the core functionnalities to simulate a virtual
35 platform. It is very low-level and is not intended to be used as
36 such but rather to serve as a basis for higher-level simulators
37 (like MSG, GRAS, SMPI, ...). It relies on a fast max min linear
40 \li MSG is a simulator built using the previous modules. It aims at
41 being realistic and is application-oriented. It is the software layer
42 of choice for building simulation with multiple scheduling agents.
44 \li GRAS (<em>not functionnal yet</em>) is an ongoing project to emulate virtual virtual platforms
45 through SURF. As a consequence a code developped using the GRAS
46 framework is able to run as well in the real-world as in the
47 simulator. If you intend to use MSG in a very intensive way
48 (e.g. for simulating a peer-to-peer environment), you may want to
51 \li SMPI (<em>not functionnal yet</em>) is an ongoing project to enable MPI code to run on top of a
52 virtual platform through SURF. It follows the same principle as
53 the ones used in GRAS but is specific to MPI applications.
55 The section \ref publications contains links to papers that provide
56 additional details on the project as well as validation and
59 The software can be downloaded from <a href="http://gcl.ucsd.edu/SimGrid/dl/">here</a>.
61 \section install Installation
64 \li <tt>./configure</tt>
65 \li <tt>make all install</tt>
67 If you are not familiar with compiling C files under UNIX and using
68 libraries, you will find some more informations in Section \ref
71 \section documentation API Documentation
73 The API of all different modules is described in \ref SimGrid_API.
75 See \ref SimGrid_examples for an introduction on the way to use these modules.
77 \section users_contributers Users / Contributers
79 \subsection contributers Contributers
81 \li Loris Marchal: wrote the new algorithm for simulation TCP
83 \li Julien Lerouge : wrote a XML parser for ENV descriptions and
84 helped for the general design during a 4 month period (march-june 2002)
86 \li Clément Menier and Marc Perache : wrote a first prototype of
87 the MSG interface during a project at ENS-Lyon (jan 2002).
88 \li Dmitrii Zagorodnov : wrote some parts of the first version
91 \subsection mailinglist User Mailing List
93 We have a <a href=https://listes.ens-lyon.fr/wws/info/simgrid2-users> mailing list for
96 \section publications Publications
98 \subsection simulation About simulation
100 \li <b>Scheduling Distributed Applications: the
101 SimGrid Simulation Framework</b>\n
102 by <em>Henri Casanova and Arnaud Legrand and Loris Marchal</em>\n
103 Proceedings of the third IEEE International Symposium
104 on Cluster Computing and the Grid (CCGrid'03)\n
105 Since the advent of distributed computer systems an active field
106 of research has been the investigation of scheduling strategies
107 for parallel applications. The common approach is to employ
108 scheduling heuristics that approximate an optimal
109 schedule. Unfortunately, it is often impossible to obtain
110 analytical results to compare the efficacy of these heuristics.
111 One possibility is to conducts large numbers of back-to-back
112 experiments on real platforms. While this is possible on
113 tightly-coupled platforms, it is infeasible on modern distributed
114 platforms (i.e. Grids) as it is labor-intensive and does not
115 enable repeatable results. The solution is to resort to
116 simulations. Simulations not only enables repeatable results but
117 also make it possible to explore wide ranges of platform and
118 application scenarios.\n
119 In this paper we present the SimGrid framework which enables the
120 simulation of distributed applications in distributed computing
121 environments for the specific purpose of developing and evaluating
122 scheduling algorithms. This paper focuses on SimGrid v2, which
123 greatly improves on the first version of the software with more
124 realistic network models and topologies. SimGrid v2 also enables
125 the simulation of distributed scheduling agents, which has become
126 critical for current scheduling research in large-scale platforms.
127 After describing and validating these features, we present a case
128 study by which we demonstrate the usefulness of SimGrid for
129 conducting scheduling research.
132 \li <b>A Network Model for Simulation of Grid Application</b>\n
133 by <em>Henri Casanova and Loris Marchal</em>\n
135 In this work we investigate network models that can be
136 potentially employed in the simulation of scheduling algorithms for
137 distributed computing applications. We seek to develop a model of TCP
138 communication which is both high-level and realistic. Previous research
139 works show that accurate and global modeling of wide-area networks, such
140 as the Internet, faces a number of challenging issues. However, some
141 global models of fairness and bandwidth-sharing exist, and can be link
142 withthe behavior of TCP. Using both previous results and simulation (with
143 NS), we attempt to understand the macroscopic behavior of
144 TCP communications. We then propose a global model of the network for the
145 Grid platform. We perform partial validation of this model in
146 simulation. The model leads to an algorithm for computing
147 bandwidth-sharing. This algorithm can then be implemented as part of Grid
148 application simulations. We provide such an implementation for the
149 SimGrid simulation toolkit.\n
150 ftp://ftp.ens-lyon.fr/pub/LIP/Rapports/RR/RR2002/RR2002-40.ps.gz
153 \li <b>MetaSimGrid : Towards realistic scheduling simulation of
154 distributed applications</b>\n
155 by <em>Arnaud Legrand and Julien Lerouge</em>\n
156 Most scheduling problems are already hard on homogeneous
157 platforms, they become quite intractable in an heterogeneous
158 framework such as a metacomputing grid. In the best cases, a
159 guaranteed heuristic can be found, but most of the time, it is
160 not possible. Real experiments or simulations are often
161 involved to test or to compare heuristics. However, on a
162 distributed heterogeneous platform, such experiments are
163 technically difficult to drive, because of the genuine
164 instability of the platform. It is almost impossible to
165 guarantee that a platform which is not dedicated to the
166 experiment, will remain exactly the same between two tests,
167 thereby forbidding any meaningful comparison. Simulations are
168 then used to replace real experiments, so as to ensure the
169 reproducibility of measured data. A key issue is the
170 possibility to run the simulations against a realistic
171 environment. The main idea of trace-based simulation is to
172 record the platform parameters today, and to simulate the
173 algorithms tomorrow, against the recorded data: even though it
174 is not the current load of the platform, it is realistic,
175 because it represents a fair summary of what happened
176 previously. A good example of a trace-based simulation tool is
177 SimGrid, a toolkit providing a set of core abstractions and
178 functionalities that can be used to easily build simulators for
179 specific application domains and/or computing environment
180 topologies. Nevertheless, SimGrid lacks a number of convenient
181 features to craft simulations of a distributed application
182 where scheduling decisions are not taken by a single
183 process. Furthermore, modeling a complex platform by hand is
184 fastidious for a few hosts and is almost impossible for a real
185 grid. This report is a survey on simulation for scheduling
186 evaluation purposes and present MetaSimGrid, a simulator built
188 ftp://ftp.ens-lyon.fr/pub/LIP/Rapports/RR/RR2002/RR2002-28.ps.gz
190 \li <b>SimGrid: A Toolkit for the Simulation of Application
192 by <em>Henri Casanova</em>\n
193 Advances in hardware and software technologies have made it
194 possible to deploy parallel applications over increasingly large
195 sets of distributed resources. Consequently, the study of
196 scheduling algorithms for such applications has been an active area
197 of research. Given the nature of most scheduling problems one must
198 resort to simulation to effectively evaluate and compare their
199 efficacy over a wide range of scenarios. It has thus become
200 necessary to simulate those algorithms for increasingly complex
201 distributed, dynamic, heterogeneous environments. In this paper we
202 present SimGrid, a simulation toolkit for the study of scheduling
203 algorithms for distributed application. This paper gives the main
204 concepts and models behind SimGrid, describes its API and
205 highlights current implementation issues. We also give some
206 experimental results and describe work that builds on SimGrid's
208 http://grail.sdsc.edu/papers/simgrid_ccgrid01.ps.gz
210 \subsection research Papers using SimGrid results
212 \li <b>Optimal algorithms for scheduling divisible workloads on
213 heterogeneous systems</b>\n
214 by <em>Olivier Beaumont and Arnaud Legrand and Yves Robert</em>\n
215 In this paper, we discuss several algorithms for scheduling
216 divisible loads on heterogeneous systems. Our main contributions
217 are (i) new optimality results for single-round algorithms and (ii)
218 the design of an asymptotically optimal multi-round algorithm. This
219 multi-round algorithm automatically performs resource selection, a
220 difficult task that was previously left to the user. Because it is
221 periodic, it is simpler to implement, and more robust to changes in
222 the speeds of processors or communication links. On the theoretical
223 side, to the best of our knowledge, this is the first published
224 result assessing the absolute performance of a multi-round
225 algorithm. On the practical side, extensive simulations reveal
226 that our multi-round algorithm outperforms existing solutions on a
227 large variety of platforms, especially when the
228 communication-to-computation ratio is not very high (the difficult
230 ftp://ftp.ens-lyon.fr/pub/LIP/Rapports/RR/RR2002/RR2002-36.ps.gz
231 \li <b>On-line Parallel Tomography</b>\n
232 by <em>Shava Smallen</em>\n
233 Masters Thesis, UCSD, May 2001
234 \li <b>Applying Scheduling and Tuning to On-line Parallel Tomography </b>\n
235 by <em>Shava Smallen, Henri Casanova, Francine Berman</em>\n
236 in Proceedings of Supercomputing 2001
237 \li <b>Heuristics for Scheduling Parameter Sweep applications in
238 Grid environments</b>\n
239 by <em>Henri Casanova, Arnaud Legrand, Dmitrii Zagorodnov and
240 Francine Berman</em>\n
241 in Proceedings of the 9th Heterogeneous Computing workshop
242 (HCW'2000), pp349-363.