Versatile Simulation of Distributed Systems
SimGrid is a toolkit providing the core functionalities for the simulation of distributed applications in heterogeneous distributed environments.
The project goal is both to facilitate research and to help improving real applications in the area of distributed and parallel systems, ranging from simple network of workstations to Computational Grids to Clouds and to supercomputers.
The goal of this practical session is to illustrate various usage of the MSG interface. To this end we will use the following simple setting:
Assume we have a (possibly large) bunch of (possibly large) data to process and which originally reside on a server (a.k.a. master). For sake of simplicity, we assume all input file require the same amount of computation. We assume the server can be helped by a (possibly large) set of worker machines. What is the best way to organize the computations ?
Although this looks like a very simple setting it raises several interesting questions:
Which algorithm should the master use to send workload?
The most obvious algorithm would be to send tasks to workers in a round-robin fashion. This is the initial code we provide you.
A less obvious but probably more efficient approach would be to set up a request mechanism where a client first ask for tasks, which allows the server to decide which request to answer and possibly to send the tasks to the fastest machines. Maybe you can think of a smarter mechanism...
How many tasks should the client ask for?
Indeed, if we set up a request mechanism so that workers only send request whenever they have no more task to process, they are likely to be poorly exploited since they will have to wait for the master to consider their request and for the input data to be transferred. A client should thus probably request a pool of tasks but if it requests too many tasks, it is likely to lead to a poor load-balancing...
How is the quality of such algorithm dependent on the platform characteristics and on the task characteristics?
Whenever the input communication time is very small compared to processing time and workers are homogeneous, it is likely that the round-robin algorithm performs very well. Would it still hold true when transfer time is not negligible and the platform is, say, a volunteer computing system ?
The network topology interconnecting the master and the workers may be quite complicated. How does such a topology impact the previous result?
When data transfers are the bottleneck, it is likely that a good modeling of the platform becomes essential. In this case, you may want to be able to account for complex platform topologies.
Do the algorithms depend on a perfect knowledge of this topology?
Should we still use a flat master worker deployment or should we use a
How is such an algorithm sensitive to external workload variation?
What if bandwidth, latency and power can vary with no warning? Shouldn't you study whether your algorithm is sensitive to such load variations?
Although an algorithm may be more efficient than another, how does it interfere with other applications?
As you can see, this very simple setting may need to evolve way beyond what you initially imagined.
Premature optimization is the root of all evil. – D.E.Knuth
Furthermore, writing your own simulator is much harder than you may imagine. This is why you should rely on an established and flexible one.
The following figure is a screenshot of triva visualizing a SimGrid simulation of two master worker applications (one in light gray and the other in dark gray) running in concurrence and showing resource usage over a long period of time.
Of course, you need to install SimGrid before taking this tutorial. Please refer to the relevant Section: Installing Simgrid.
A lot of information on how to install and use Simgrid are provided by the online documentation and by several tutorials:
Several tools can be used to visualize the result of SimGrid simulations and get a better understanding of simulations.
Under Debian or Ubuntu, this is really easy with apt-get, while you may have to install from the source on other systems. Check the documentation of each software for more details.
The corresponding source files can be obtained online on GitLab. If you find the right button on the top right of the interface, you can download the whole directory in one archive file. If you wish, you can find other platform file in this GitLab directory.
As you can see, there is already a little Makefile that compiles everything for you. If you struggle with the compilation, then you should double check your SimGrid installation. On need, please refer to the Troubleshooting your project setup section.
Once the tiny example has been compiled and it can be easily run as follows:
For a more "fancy" output, you can use simgrid-colorizer.
If you installed SimGrid to a non-standard path, you may have to specify the full path to simgrid-colorizer on the above line, such as
/opt/simgrid/bin/simgrid-colorizer. If you did not install it at all, you can find it in <simgrid_root_directory>/bin/colorize.
For a really fancy output, you should use viva/triva:
For a more classical Gantt-Chart visualization, you can produce a Paje trace:
Alternatively, you can use vite.
In the previous example, the deployment file
deployment0.xml is tightly connected to the platform file
platform.xml and a worker process is launched on each host:
This is ok as the platform is rather small but will be painful when using larger platforms. Instead, modify the simulator
masterworker1.c so that the master launches a worker process on all the other machines at startup. The new deployment file
deployment1.xml should thus now simply be:
To this end you may need the following MSG functions (click on the links to see their descriptions):
data field of the MSG_process_create can be used to pass a channel name that will be private between master and workers (e.g.,
master_name:worker_name). Adding the
master_name in the channel name will allow to easily have several masters and a worker per master on each machine. To this end, you may need to use the following functions:
If you are not too familiar with string manipulation in C, you may want to use the following functions (see the C reference for details):
In the current version, the number of tasks is defined through the worker arguments. Hence, tasks are created at the very beginning of the simulation. Instead, create tasks as needed and provide a time limit indicating when it stops sending tasks. To this end, you will obviously need to know what time it is:
Otherwise, a quite effective way of terminating the simulation would be to use some of the following functions:
Anyway, the new deployment
deployment2.xml file should thus look like this:
It may also be a good idea to transform most of the
XBT_DEBUG (e.g., keep the information on the total number of tasks processed). These debug messages can be activated as follows:
SimGrid can trace all resource consumption and the outcome can be displayed with viva as illustrated in the section intro_setup. However, when several masters are deployed, it is hard to understand what happens.
So let's use categories to track more precisely who does what and when:
The outcome can then be visualized as follows:
Right now, you should realize that nothing is behaving like you expect. Most workers are idle even though input data are ridiculous and there are several masters deployed on the platform. Using a Gantt-chart visualization may help:
OK, so it should now be obvious that round robin is actually very bad.
Instead of a round-robin scheduling, let's implement a first-come first-served mechanism. To this end, workers need to send a tiny request first. A possible way to implement such a request with MSG is to send on a specific channel (e.g., the name of the master name) a task with payload 0 and whose attached data is the worker name. This way, the master can keep track of which workers are idle and willing to work.
To know whether it has pending requests, the master can use the following [function][fn:7]:
If so, it should get the request and push the corresponding host into a dynar so that they can later be retrieved when sending a real [task][fn:7].
As you will soon realize, with such simple mechanisms, simple deadlocks will soon appear. They can easily be removed with a simple polling mechanism, hence the need for the following [function][fn:7]:
As you should quickly realize, on the simple previous example, it will double the throughput of the platform but will be quite ineffective when input size of the tasks is not negligible anymore.
From this, many things can easily be added. For example, you could:
SimGrid offers a rather powerful platform modeling mechanism. The
src/examples/platforms/ repository comprises a variety of platforms ranging from simple to elaborate. Associated to a good visualization tool to ensure your simulation is meaningful, they can allow you to study to which extent your algorithm scales...
What is the largest number of tasks requiring 50e6 flops and 1e5 bytes that you manage to distribute and process in one hour on
g5k.xml (you should use