6 SimGrid is a framework to simulate distributed computer systems.
8 It can be used to either assess abstract algorithms or to profile and
9 debug real distributed applications. SimGrid enables studies in the
10 domains of (data-)Grids, IaaS Clouds, Clusters, High Performance
11 Computing, Volunteer Computing, and Peer-to-Peer systems.
13 Technically speaking, SimGrid is a library. It is neither a graphical
14 interface nor a command-line simulator running user scripts. The
15 interaction with SimGrid is done by writing programs with the exposed
16 functions to build your own simulator.
18 SimGrid is a Free Software distributed under the LGPLv3 license. You are
19 thus welcome to use it as you wish or even to modify and distribute
20 your version (provided that your version is as free as ours). It also
21 means that SimGrid is developed by a vivid community of users and
22 developers. We hope that you will come and join us!
24 SimGrid is the result of almost 20 years of research from several
25 groups, both in France and in the U.S.A. It benefited of many funding
26 from various research bodies, including the ANR, Inria, CNRS,
27 University of Lorraine, University of Hawai'i at Manoa, ENS Rennes, and
28 many others. Many thanks to our generous sponsors!
30 SimGrid is a powerful tool, but its learning curve can be rather
31 steep. This manual will hopefully help and guide you to the features
32 you want to use. Please report any issue that you see in this manual,
33 including typos or unclear elements. You can even propose changes by
34 clicking on the "Edit on GitLab" button at the top of every page.
36 Typical Study based on SimGrid
37 ------------------------------
41 <object data="graphical-toc.svg" width="100%" type="image/svg+xml"></object>
44 Any SimGrid study entails the following components:
46 - The studied **Application**. This can be either a distributed
47 algorithm described in our simple APIs, or a full featured real
48 parallel application using for example the MPI interface
49 :ref:`(more info) <application>`.
51 - The **Simulated Platform**. This is a description of a given
52 distributed system (machines, links, disks, clusters, etc). Most of
53 the platform files are written in XML althrough a Lua interface is
54 under development. SimGrid makes it easy to augment the Simulated
55 Platform with a Dynamic Scenario where for example the links are
56 slowed down (because of external usage) or the machines fail. You
57 have even support to specify the applicative workload that you want
58 to feed to your application
59 :ref:`(more info) <platform>`.
61 - The application's **Deployment Description**. In SimGrid
62 terminology, the application is an inert set of source files and
63 binaries. To make it run, you have to describe how your application
64 should be deployed on the simulated platform. You need to specify
65 which process is mapped on which machine, along with their parameters
66 :ref:`(more info) <scenario>`.
68 - The **Platform Models**. They describe how the simulated platform
69 reacts to the actions of the application. For example, they compute
70 the time taken by a given communication on the simulated platform.
71 These models are already included in SimGrid, and you only need to
72 pick one and maybe tweak its configuration to get your results
73 :ref:`(more info) <models>`.
75 These components are put together to run a **simulation**, that is an
76 experiment or a probe. The result of one or many simulation provides
77 an **outcome** (logs, visualization, or statistical analysis) that help
78 answering the **question** targeted by this study.
80 Here are some questions on which SimGrid is particularly relevant:
82 - **Compare an Application to another**. This is the classical use
83 case for scientists, who use SimGrid to test how the solution that
84 they contribute to compares to the existing solutions from the
87 - **Design the best [Simulated] Platform for a given Application.**
88 Tweaking the platform file is much easier than building a new real
89 platform for testing purpose. SimGrid also allows for the co-design
90 of the platform and the application by modifying both of them.
92 - **Debug Real Applications**. With real systems, is sometimes
93 difficult to reproduce the exact run leading to the bug that you
94 are tracking. With SimGrid, you are *clairvoyant* about your
95 *reproducible experiments*: you can explore every part of the
96 system, and your probe will not change the simulated state. It also
97 makes it easy to mock some parts of the real system that are not
100 Depending on the context, you may see some parts of this process as
101 less important, but you should pay close attention if you want to be
102 confident in the results coming out of your simulations. In
103 particular, you should not blindly trust your results but always
104 strive to double-check them. Likewise, you should question the realism
105 of your input configuration, and we even encourage you to doubt (and
106 check) the provided performance models.
108 To ease such questioning, you really should logically separate these
109 parts in your experimental setup. It is seen as a very bad practice to
110 merge the application, the platform, and the deployment all together.
111 SimGrid is versatile and your mileage may vary, but you should start
112 with your Application specified as a C++ or Java program, using one of
113 the provided XML platform file, and with your deployment in a separate
116 SimGrid Execution Modes
117 -----------------------
119 Depending on the intended study, SimGrid can be run in several execution modes.
121 **Simulation Mode**. This is the most common execution mode, where you want
122 to study how your application behaves on the simulated platform under
123 the experimental scenario.
125 In this mode, SimGrid can provide information about the time taken by
126 your application, the amount of energy dissipated by the platform to
127 run your application, and the detailed usage of each resource.
129 **Model-Checking Mode**. This can be seen as a sort of exhaustive
130 testing mode, where every possible outcome of your application is
131 explored. In some sense, this mode tests your application for all
132 possible platforms that you could imagine (and more).
134 You just provide the application and its deployment (amount of
135 processes and parameters), and the model-checker will literally
136 explore all possible outcomes by testing all possible message
137 interleavings: if at some point a given process can either receive the
138 message A first or the message B depending on the platform
139 characteristics, the model-checker will explore the scenario where A
140 arrives first, and then rewind to the same point to explore the
141 scenario where B arrives first.
143 This is a very powerful mode, where you can evaluate the correction of
144 your application. It can verify either **safety properties** (asserts)
145 or **liveless properties** stating for example that if a given event
146 occurs, then another given event will occur in a finite amount of
147 steps. This mode is not only usable with the abstract algorithms
148 developed on top of the SimGrid APIs, but also with real MPI
149 applications (to some extent).
151 The main limit of Model Checking lays in the huge amount of scenarios
152 to explore. SimGrid tries to explore only non-redundant scenarios
153 thanks to classical reduction techniques (such as DPOR and stateful
154 exploration) but the exploration may well never finish if you don't
155 carefully adapt your application to this mode.
157 A classical trap is that the Model Checker can only verify whether
158 your application fits the provided properties, which is useless if you
159 have a bug in your property. Remember also that one way for your
160 application to never violate a given assert is to not start at all
161 because of a stupid bug.
163 Another limit of this mode is that it does not use the performance
164 models of the simulation mode. Time becomes discrete: You can say for
165 example that the application took 42 steps to run, but there is no way
166 to know how much time it took or the amount of watts that were dissipated.
168 Finally, the model checker only explores the interleavings of
169 computations and communications. Other factors such as thread
170 execution interleaving are not considered by the SimGrid model
173 The model checker may well miss existing issues, as it computes the
174 possible outcomes *from a given initial situation*. There is no way to
175 prove the correction of your application in all generality with this
178 **Benchmark Recording Mode**. During debug sessions, continuous
179 integration testing, and other similar use cases, you are often only
180 interested in the control flow. If your application apply filters to
181 huge images split in small blocks, the filtered image is probably not
182 what you are interested in. You are probably looking for a way to run
183 each computation kernel only once, save on disk the time it takes and
184 some other metadata. This code block can then be skipped in simulation
185 and replaced by a synthetic block using the cached information. The
186 simulated platform will take this block into account without requesting
187 the real hosting machine to benchmark it.
192 This framework is by no means the perfect holly grail able to solve
193 every problem on earth.
195 **SimGrid scope is limited to distributed systems.** Real-time
196 multi-threaded systems are out of scope. You could probably tweak
197 SimGrid for such studies (or the framework could possibly be extended
198 in this direction), but another framework specifically targeting such a
199 use case would probably be more suited.
201 **There is currently no support for wireless networks**.
202 The framework could certainly be improved in this direction, but this
203 still has to be done.
205 **There is no perfect model, only models adapted to your study.**
206 The SimGrid models target fast and large studies yet requesting
207 realistic results. In particular, our models abstract away parameters
208 and phenomena that are often irrelevant to the realism in our
211 SimGrid is simply not intended to any study that would mandate the
212 abstracted phenomenon. Here are some **studies that you should not do
215 - Studying the effect of L3 vs. L2 cache effects on your application
216 - Comparing kernel schedulers and policies
217 - Comparing variants of TCP
218 - Exploring pathological cases where TCP breaks down, resulting in
220 - Studying security aspects of your application, in presence of
223 SimGrid Success Stories
224 -----------------------
226 SimGrid was cited in over 1,500 scientific papers (according to Google
228 `over 200 publications <https://simgrid.org/Usages.html>`_
229 (written by about 300 individuals) use SimGrid as a scientific
230 instrument to conduct their experimental evaluation. These
231 numbers do not include the articles contributing to SimGrid.
232 This instrument was used in many research communities, such as
233 `High-Performance Computing <https://hal.inria.fr/inria-00580599/>`_,
234 `Cloud Computing <http://dx.doi.org/10.1109/CLOUD.2015.125>`_,
235 `Workflow Scheduling <http://dl.acm.org/citation.cfm?id=2310096.2310195>`_,
236 `Big Data <https://hal.inria.fr/hal-01199200/>`_ and
237 `MapReduce <http://dx.doi.org/10.1109/WSCAD-SSC.2012.18>`_,
238 `Data Grid <http://ieeexplore.ieee.org/document/7515695/>`_,
239 `Volunteer Computing <http://www.sciencedirect.com/science/article/pii/S1569190X17301028>`_,
240 `Peer-to-Peer Computing <https://hal.archives-ouvertes.fr/hal-01152469/>`_,
241 `Network Architecture <http://dx.doi.org/10.1109/TPDS.2016.2613043>`_,
242 `Fog Computing <http://ieeexplore.ieee.org/document/7946412/>`_, or
243 `Batch Scheduling <https://hal.archives-ouvertes.fr/hal-01333471>`_
244 `(more info) <https://simgrid.org/Usages.html>`_.
246 If your platform description is accurate enough (see
247 `here <http://hal.inria.fr/hal-00907887>`_ or
248 `there <https://hal.inria.fr/hal-01523608>`_),
249 SimGrid can provide high-quality performance predictions. For example,
250 we determined the speedup achieved by the Tibidabo ARM-based
251 cluster before its construction
252 (`paper <http://hal.inria.fr/hal-00919507>`_). In this case,
253 some differences between the prediction and the real timings were due to
254 misconfiguration or other problems with the real platform. To some extent,
255 SimGrid could even be used to debug the real platform :)
257 SimGrid is also used to debug, improve, and tune several large
259 `BigDFT <http://bigdft.org>`_ (a massively parallel code
260 computing the electronic structure of chemical elements developped by
261 the CEA), `StarPU <http://starpu.gforge.inria.fr/>`_ (a
262 Unified Runtime System for Heterogeneous Multicore Architectures
263 developped by Inria Bordeaux) and
264 `TomP2P <https://tomp2p.net/dev/simgrid/>`_ (a high performance
265 key-value pair storage library developed at University of Zurich).
266 Some of these applications enjoy large user communities themselves.
268 .. LocalWords: SimGrid
