Typical Study based on SimGrid¶
Any SimGrid study entails the following components:
- The studied Application. This can be either a distributed algorithm described in our simple APIs, or a full featured real parallel application using for example the MPI interface (more info).
- The Simulated Platform. This is a description of a given distributed system (machines, links, disks, clusters, etc). Most of the platform files are written in XML althrough a Lua interface is under development. SimGrid makes it easy to augment the Simulated Platform with a Dynamic Scenario where for example the links are slowed down (because of external usage) or the machines fail. You have even support to specify the applicative workload that you want to feed to your application (more info).
- The application’s Deployment Description. In SimGrid terminology, the application is an inert set of source files and binaries. To make it run, you have to describe how your application should be deployed on the simulated platform. You need to specify which process is mapped on which machine, along with their parameters (more info).
- The Platform Models. They describe how the simulated platform reacts to the actions of the application. For example, they compute the time taken by a given communication on the simulated platform. These models are already included in SimGrid, and you only need to pick one and maybe tweak its configuration to get your results (more info).
These components are put together to run a simulation, that is an experiment or a probe. The result of one or many simulation provides an outcome (logs, visualization, or statistical analysis) that help answering the question targeted by this study.
Here are some questions on which SimGrid is particularly relevant:
- Compare an Application to another. This is the classical use case for scientists, who use SimGrid to test how the solution that they contribute to compares to the existing solutions from the literature.
- Design the best [Simulated] Platform for a given Application. Tweaking the platform file is much easier than building a new real platform for testing purpose. SimGrid also allows for the co-design of the platform and the application by modifying both of them.
- Debug Real Applications. With real systems, is sometimes difficult to reproduce the exact run leading to the bug that you are tracking. With SimGrid, you are clairvoyant about your reproducible experiments: you can explore every part of the system, and your probe will not change the simulated state. It also makes it easy to mock some parts of the real system that are not under study.
Depending on the context, you may see some parts of this process as less important, but you should pay close attention if you want to be confident in the results coming out of your simulations. In particular, you should not blindly trust your results but always strive to double-check them. Likewise, you should question the realism of your input configuration, and we even encourage you to doubt (and check) the provided performance models.
To ease such questioning, you really should logically separate these parts in your experimental setup. It is seen as a very bad practice to merge the application, the platform, and the deployment all together. SimGrid is versatile and your mileage may vary, but you should start with your Application specified as a C++ or Java program, using one of the provided XML platform file, and with your deployment in a separate XML file.
SimGrid Execution Modes¶
Depending on the intended study, SimGrid can be run in several execution modes.
Simulation Mode. This is the most common execution mode, where you want to study how your application behaves on the simulated platform under the experimental scenario.
In this mode, SimGrid can provide information about the time taken by your application, the amount of energy dissipated by the platform to run your application, and the detailed usage of each resource.
Model-Checking Mode. This can be seen as a sort of exhaustive testing mode, where every possible outcome of your application is explored. In some sense, this mode tests your application for all possible platforms that you could imagine (and more).
You just provide the application and its deployment (amount of processes and parameters), and the model-checker will literally explore all possible outcomes by testing all possible message interleavings: if at some point a given process can either receive the message A first or the message B depending on the platform characteristics, the model-checker will explore the scenario where A arrives first, and then rewind to the same point to explore the scenario where B arrives first.
This is a very powerful mode, where you can evaluate the correction of your application. It can verify either safety properties (asserts) or liveless properties stating for example that if a given event occurs, then another given event will occur in a finite amount of steps. This mode is not only usable with the abstract algorithms developed on top of the SimGrid APIs, but also with real MPI applications (to some extent).
The main limit of Model Checking lays in the huge amount of scenarios to explore. SimGrid tries to explore only non-redundant scenarios thanks to classical reduction techniques (such as DPOR and stateful exploration) but the exploration may well never finish if you don’t carefully adapt your application to this mode.
A classical trap is that the Model Checker can only verify whether your application fits the provided properties, which is useless if you have a bug in your property. Remember also that one way for your application to never violate a given assert is to not start at all because of a stupid bug.
Another limit of this mode is that it does not use the performance models of the simulation mode. Time becomes discrete: You can say for example that the application took 42 steps to run, but there is no way to know how much time it took or the amount of watts that were dissipated.
Finally, the model checker only explores the interleavings of computations and communications. Other factors such as thread execution interleaving are not considered by the SimGrid model checker.
The model checker may well miss existing issues, as it computes the possible outcomes from a given initial situation. There is no way to prove the correction of your application in all generality with this tool.
Benchmark Recording Mode. During debug sessions, continuous integration testing, and other similar use cases, you are often only interested in the control flow. If your application apply filters to huge images split in small blocks, the filtered image is probably not what you are interested in. You are probably looking for a way to run each computation kernel only once, save on disk the time it takes and some other metadata. This code block can then be skipped in simulation and replaced by a synthetic block using the cached information. The simulated platform will take this block into account without requesting the real hosting machine to benchmark it.
This framework is by no means the perfect holly grail able to solve every problem on earth.
SimGrid scope is limited to distributed systems. Real-time multi-threaded systems are out of scope. You could probably tweak SimGrid for such studies (or the framework could possibly be extended in this direction), but another framework specifically targeting such a use case would probably be more suited.
There is currently no support for wireless networks. The framework could certainly be improved in this direction, but this still has to be done.
There is no perfect model, only models adapted to your study. The SimGrid models target fast and large studies yet requesting realistic results. In particular, our models abstract away parameters and phenomena that are often irrelevant to the realism in our context.
SimGrid is simply not intended to any study that would mandate the abstracted phenomenon. Here are some studies that you should not do with SimGrid:
- Studying the effect of L3 vs. L2 cache effects on your application
- Comparing kernel schedulers and policies
- Comparing variants of TCP
- Exploring pathological cases where TCP breaks down, resulting in abnormal executions.
- Studying security aspects of your application, in presence of malicious agents.
SimGrid Success Stories¶
SimGrid was cited in over 1,500 scientific papers (according to Google Scholar). Among them over 200 publications (written by about 300 individuals) use SimGrid as a scientific instrument to conduct their experimental evaluation. These numbers do not include the articles contributing to SimGrid. This instrument was used in many research communities, such as High-Performance Computing, Cloud Computing, Workflow Scheduling, Big Data and MapReduce, Data Grid, Volunteer Computing, Peer-to-Peer Computing, Network Architecture, Fog Computing, or Batch Scheduling (more info).
If your platform description is accurate enough (see here or there), SimGrid can provide high-quality performance predictions. For example, we determined the speedup achieved by the Tibidabo ARM-based cluster before its construction (paper). In this case, some differences between the prediction and the real timings were due to misconfiguration or other problems with the real platform. To some extent, SimGrid could even be used to debug the real platform :)
SimGrid is also used to debug, improve, and tune several large applications. BigDFT (a massively parallel code computing the electronic structure of chemical elements developped by the CEA), StarPU (a Unified Runtime System for Heterogeneous Multicore Architectures developped by Inria Bordeaux) and TomP2P (a high performance key-value pair storage library developed at University of Zurich). Some of these applications enjoy large user communities themselves.