6 Typical Study based on SimGrid
7 ------------------------------
11 <object data="graphical-toc.svg" width="100%" type="image/svg+xml"></object>
14 Any SimGrid study entails the following components:
16 - The studied **Application**. This can be either a distributed
17 algorithm described in our simple APIs, or a full featured real
18 parallel application using for example the MPI interface
19 :ref:`(more info) <application>`.
21 - The **Simulated Platform**. This is a description of a given
22 distributed system (machines, links, disks, clusters, etc). Most of
23 the platform files are written in XML althrough a Lua interface is
24 under development. SimGrid makes it easy to augment the Simulated
25 Platform with a Dynamic Scenario where for example the links are
26 slowed down (because of external usage) or the machines fail. You
27 have even support to specify the applicative workload that you want
28 to feed to your application
29 :ref:`(more info) <platform>`.
31 - The application's **Deployment Description**. In SimGrid
32 terminology, the application is an inert set of source files and
33 binaries. To make it run, you have to describe how your application
34 should be deployed on the simulated platform. You need to specify
35 which process is mapped on which machine, along with their parameters
36 :ref:`(more info) <scenario>`.
38 - The **Platform Models**. They describe how the simulated platform
39 reacts to the actions of the application. For example, they compute
40 the time taken by a given communication on the simulated platform.
41 These models are already included in SimGrid, and you only need to
42 pick one and maybe tweak its configuration to get your results
43 :ref:`(more info) <models>`.
45 These components are put together to run a **simulation**, that is an
46 experiment or a probe. The result of one or many simulation provides
47 an **outcome** (logs, visualization, or statistical analysis) that help
48 answering the **question** targeted by this study.
50 Here are some questions on which SimGrid is particularly relevant:
52 - **Compare an Application to another**. This is the classical use
53 case for scientists, who use SimGrid to test how the solution that
54 they contribute to compares to the existing solutions from the
57 - **Design the best [Simulated] Platform for a given Application.**
58 Tweaking the platform file is much easier than building a new real
59 platform for testing purpose. SimGrid also allows for the co-design
60 of the platform and the application by modifying both of them.
62 - **Debug Real Applications**. With real systems, is sometimes
63 difficult to reproduce the exact run leading to the bug that you
64 are tracking. With SimGrid, you are *clairvoyant* about your
65 *reproducible experiments*: you can explore every part of the
66 system, and your probe will not change the simulated state. It also
67 makes it easy to mock some parts of the real system that are not
70 Depending on the context, you may see some parts of this process as
71 less important, but you should pay close attention if you want to be
72 confident in the results coming out of your simulations. In
73 particular, you should not blindly trust your results but always
74 strive to double-check them. Likewise, you should question the realism
75 of your input configuration, and we even encourage you to doubt (and
76 check) the provided performance models.
78 To ease such questioning, you really should logically separate these
79 parts in your experimental setup. It is seen as a very bad practice to
80 merge the application, the platform, and the deployment all together.
81 SimGrid is versatile and your mileage may vary, but you should start
82 with your Application specified as a C++ or Java program, using one of
83 the provided XML platform file, and with your deployment in a separate
86 SimGrid Execution Modes
87 -----------------------
89 Depending on the intended study, SimGrid can be run in several execution modes.
91 **Simulation Mode**. This is the most common execution mode, where you want
92 to study how your application behaves on the simulated platform under
93 the experimental scenario.
95 In this mode, SimGrid can provide information about the time taken by
96 your application, the amount of energy dissipated by the platform to
97 run your application, and the detailed usage of each resource.
99 **Model-Checking Mode**. This can be seen as a sort of exhaustive
100 testing mode, where every possible outcome of your application is
101 explored. In some sense, this mode tests your application for all
102 possible platforms that you could imagine (and more).
104 You just provide the application and its deployment (amount of
105 processes and parameters), and the model-checker will literally
106 explore all possible outcomes by testing all possible message
107 interleavings: if at some point a given process can either receive the
108 message A first or the message B depending on the platform
109 characteristics, the model-checker will explore the scenario where A
110 arrives first, and then rewind to the same point to explore the
111 scenario where B arrives first.
113 This is a very powerful mode, where you can evaluate the correction of
114 your application. It can verify either **safety properties** (asserts)
115 or **liveless properties** stating for example that if a given event
116 occurs, then another given event will occur in a finite amount of
117 steps. This mode is not only usable with the abstract algorithms
118 developed on top of the SimGrid APIs, but also with real MPI
119 applications (to some extent).
121 The main limit of Model Checking lays in the huge amount of scenarios
122 to explore. SimGrid tries to explore only non-redundant scenarios
123 thanks to classical reduction techniques (such as DPOR and stateful
124 exploration) but the exploration may well never finish if you don't
125 carefully adapt your application to this mode.
127 A classical trap is that the Model Checker can only verify whether
128 your application fits the provided properties, which is useless if you
129 have a bug in your property. Remember also that one way for your
130 application to never violate a given assert is to not start at all
131 because of a stupid bug.
133 Another limit of this mode is that it does not use the performance
134 models of the simulation mode. Time becomes discrete: You can say for
135 example that the application took 42 steps to run, but there is no way
136 to know how much time it took or the amount of watts that were dissipated.
138 Finally, the model checker only explores the interleavings of
139 computations and communications. Other factors such as thread
140 execution interleaving are not considered by the SimGrid model
143 The model checker may well miss existing issues, as it computes the
144 possible outcomes *from a given initial situation*. There is no way to
145 prove the correction of your application in all generality with this
148 **Benchmark Recording Mode**. During debug sessions, continuous
149 integration testing, and other similar use cases, you are often only
150 interested in the control flow. If your application apply filters to
151 huge images split in small blocks, the filtered image is probably not
152 what you are interested in. You are probably looking for a way to run
153 each computation kernel only once, save on disk the time it takes and
154 some other metadata. This code block can then be skipped in simulation
155 and replaced by a synthetic block using the cached information. The
156 simulated platform will take this block into account without requesting
157 the real hosting machine to benchmark it.
162 This framework is by no means the perfect holly grail able to solve
163 every problem on earth.
165 **SimGrid scope is limited to distributed systems.** Real-time
166 multi-threaded systems are out of scope. You could probably tweak
167 SimGrid for such studies (or the framework could possibly be extended
168 in this direction), but another framework specifically targeting such a
169 use case would probably be more suited.
171 **There is currently no support for wireless networks**.
172 The framework could certainly be improved in this direction, but this
173 still has to be done.
175 **There is no perfect model, only models adapted to your study.**
176 The SimGrid models target fast and large studies yet requesting
177 realistic results. In particular, our models abstract away parameters
178 and phenomena that are often irrelevant to the realism in our
181 SimGrid is simply not intended to any study that would mandate the
182 abstracted phenomenon. Here are some **studies that you should not do
185 - Studying the effect of L3 vs. L2 cache effects on your application
186 - Comparing kernel schedulers and policies
187 - Comparing variants of TCP
188 - Exploring pathological cases where TCP breaks down, resulting in
190 - Studying security aspects of your application, in presence of
193 SimGrid Success Stories
194 -----------------------
196 SimGrid was cited in over 1,500 scientific papers (according to Google
198 `over 200 publications <https://simgrid.org/Usages.html>`_
199 (written by about 300 individuals) use SimGrid as a scientific
200 instrument to conduct their experimental evaluation. These
201 numbers do not include the articles contributing to SimGrid.
202 This instrument was used in many research communities, such as
203 `High-Performance Computing <https://hal.inria.fr/inria-00580599/>`_,
204 `Cloud Computing <http://dx.doi.org/10.1109/CLOUD.2015.125>`_,
205 `Workflow Scheduling <http://dl.acm.org/citation.cfm?id=2310096.2310195>`_,
206 `Big Data <https://hal.inria.fr/hal-01199200/>`_ and
207 `MapReduce <http://dx.doi.org/10.1109/WSCAD-SSC.2012.18>`_,
208 `Data Grid <http://ieeexplore.ieee.org/document/7515695/>`_,
209 `Volunteer Computing <http://www.sciencedirect.com/science/article/pii/S1569190X17301028>`_,
210 `Peer-to-Peer Computing <https://hal.archives-ouvertes.fr/hal-01152469/>`_,
211 `Network Architecture <http://dx.doi.org/10.1109/TPDS.2016.2613043>`_,
212 `Fog Computing <http://ieeexplore.ieee.org/document/7946412/>`_, or
213 `Batch Scheduling <https://hal.archives-ouvertes.fr/hal-01333471>`_
214 `(more info) <https://simgrid.org/Usages.html>`_.
216 If your platform description is accurate enough (see
217 `here <http://hal.inria.fr/hal-00907887>`_ or
218 `there <https://hal.inria.fr/hal-01523608>`_),
219 SimGrid can provide high-quality performance predictions. For example,
220 we determined the speedup achieved by the Tibidabo ARM-based
221 cluster before its construction
222 (`paper <http://hal.inria.fr/hal-00919507>`_). In this case,
223 some differences between the prediction and the real timings were due to
224 misconfiguration or other problems with the real platform. To some extent,
225 SimGrid could even be used to debug the real platform :)
227 SimGrid is also used to debug, improve, and tune several large
229 `BigDFT <http://bigdft.org>`_ (a massively parallel code
230 computing the electronic structure of chemical elements developped by
231 the CEA), `StarPU <http://starpu.gforge.inria.fr/>`_ (a
232 Unified Runtime System for Heterogeneous Multicore Architectures
233 developped by Inria Bordeaux) and
234 `TomP2P <https://tomp2p.net/dev/simgrid/>`_ (a high performance
235 key-value pair storage library developed at University of Zurich).
236 Some of these applications enjoy large user communities themselves.
238 .. LocalWords: SimGrid
