Configuring SimGrid



A number of options can be given at runtime to change the default SimGrid behavior. For a complete list of all configuration options accepted by the SimGrid version used in your simulator, simply pass the –help configuration flag to your program. If some of the options are not documented on this page, this is a bug that you should please report so that we can fix it. Note that some of the options presented here may not be available in your simulators, depending on the compile-time options that you used.

Setting Configuration Items

There is several way to pass configuration options to the simulators. The most common way is to use the --cfg command line argument. For example, to set the item Item to the value Value, simply type the following on the command-line:

$ my_simulator --cfg=Item:Value (other arguments)

Several --cfg command line arguments can naturally be used. If you need to include spaces in the argument, don’t forget to quote the argument. You can even escape the included quotes (write @' for ' if you have your argument between simple quotes).

Another solution is to use the <config> tag in the platform file. The only restriction is that this tag must occur before the first platform element (be it <zone>, <cluster>, <peer> or whatever). The <config> tag takes an id attribute, but it is currently ignored so you don’t really need to pass it. The important part is that within that tag, you can pass one or several <prop> tags to specify the configuration to use. For example, setting Item to Value can be done by adding the following to the beginning of your platform file:

<config>
  <prop id="Item" value="Value"/>
</config>

A last solution is to pass your configuration directly in your program with simgrid::s4u::Engine::set_config().

#include <simgrid/s4u.hpp>

int main(int argc, char *argv[]) {
  simgrid::s4u::Engine e(&argc, argv);

  simgrid::s4u::Engine::set_config("Item:Value");

  // Rest of your code
}

Existing Configuration Items

Note

The full list can be retrieved by passing --help and --help-cfg to an executable that uses SimGrid. Try passing help as a value to get the list of values accepted by a given option. For example, --cfg=plugin:help will give you the list of plugins available in your installation of SimGrid.

Configuring the Platform Models

Choosing the Platform Models

SimGrid comes with several network, CPU and disk models built in, and you can change the used model at runtime by changing the passed configuration. The three main configuration items are given below. For each of these items, passing the special help value gives you a short description of all possible values (for example, --cfg=network/model:help will present all provided network models). Also, --help-models should provide information about all models for all existing resources.

  • network/model: specify the used network model. Possible values:

    • LV08 (default one): Realistic network analytic model (slow-start modeled by multiplying latency by 13.01, bandwidth by .97; bottleneck sharing uses a payload of S=20537 for evaluating RTT). Presented in the relevant section.

    • Constant: Simplistic network model where all communication take a constant time (one second). This model provides the lowest realism, but is (marginally) faster.

    • SMPI: Realistic network model specifically tailored for HPC settings (accurate modeling of slow start with correction factors on three intervals: < 1KiB, < 64 KiB, >= 64 KiB). This model can be further configured.

    • IB: Realistic network model specifically tailored for HPC settings with InfiniBand networks (accurate modeling contention behavior, based on the model explained in this PhD work. This model can be further configured.

    • CM02: Legacy network analytic model. Very similar to LV08, but without corrective factors. The timings of small messages are thus poorly modeled. Presented in the relevant section.

    • ns-3 (only available if you compiled SimGrid accordingly): Use the packet-level network simulators as network models (see ns-3 as a SimGrid model). This model can be further configured.

  • cpu/model: specify the used CPU model. We have only one model for now:

    • Cas01: Simplistic CPU model (time=size/speed)

  • host/model: we have two such models for now.

    • default: Default host model. It simply uses the otherwise configured models for cpu, disk and network (i.e. CPU:Cas01, disk:S19 and network:LV08 by default)

    • ptask_L07: This model is mandatory if you plan to use parallel tasks (and useless otherwise). ptasks are intended to model the moldable tasks of the grid scheduling literature. A specific host model is necessary because each such activity has a both compute and communicate components, so the CPU and network models must be mixed together.

  • storage/model: specify the used storage model. Only one model is provided so far.

  • vm/model: specify the model for virtual machines. Only one model is provided so far.

Solver

The different models rely on a linear inequalities solver to share the underlying resources. SimGrid allows you to change the solver, but be cautious, don’t change it unless you are 100% sure.

  • items cpu/solver, network/solver, disk/solver and host/solver allow you to change the solver for each model:

    • maxmin: The default solver for all models except ptask. Provides a max-min fairness allocation.

    • fairbottleneck: The default solver for ptasks. Extends max-min to allow heterogeneous resources.

    • bmf: More realistic solver for heterogeneous resource sharing. Implements BMF (Bottleneck max fairness) fairness. To be used with parallel tasks instead of fair-bottleneck.

Optimization Level

The network and CPU models that are based on linear inequalities solver (that is, all our analytical models) accept specific optimization configurations.

  • items network/optim and cpu/optim (both default to ‘Lazy’):

    • Lazy: Lazy action management (partial invalidation in lmm + heap in action remaining).

    • TI: Trace integration. Highly optimized mode when using availability traces (only available for the Cas01 CPU model for now).

    • Full: Full update of remaining and variables. Slow but may be useful when debugging.

  • items network/maxmin-selective-update and cpu/maxmin-selective-update: configure whether the underlying should be lazily updated or not. It should have no impact on the computed timings, but should speed up the computation.
    It is still possible to disable this feature because it can reveal counter-productive in very specific scenarios where the interaction level is high. In particular, if all your communication share a given backbone link, you should disable it: without it, a simple regular loop is used to update each communication. With it, each of them is still updated (because of the dependency induced by the backbone), but through a complicated and slow pattern that follows the actual dependencies.

Numerical Precision

Option precision/timing Default: 1e-9 (in seconds)
Option precision/work-amount Default: 1e-5 (in flops or bytes)
Option bmf/precision Default: 1e-12 (no unit)

The analytical models handle a lot of floating point values. It is possible to change the epsilon used to update and compare them through this configuration item. Changing it may speedup the simulation by discarding very small actions, at the price of a reduced numerical precision. You can modify separately the precision used to manipulate timings (in seconds) and the one used to manipulate amounts of work (in flops or bytes).

Concurrency Limit

Option maxmin/concurrency-limit Default: -1 (no limit)

The maximum number of variables per resource can be tuned through this option. You can have as many simultaneous actions per resources as you want. If your simulation presents a very high level of concurrency, it may help to use e.g. 100 as a value here. It means that at most 100 actions can consume a resource at a given time. The extraneous actions are queued and wait until the amount of concurrency of the considered resource lowers under the given boundary.

Such limitations help both to the simulation speed and simulation accuracy on highly constrained scenarios, but the simulation speed suffers of this setting on regular (less constrained) scenarios so it is off by default.

BMF settings

Option bmf/max-iterations Default: 1000

It may happen in some settings that the BMF solver fails to converge to a solution, so there is a hard limit on the amount of iteration count to avoid infinite loops.

Configuring the Network Model

Maximal TCP Window Size

Option network/TCP-gamma Default: 4194304

The analytical models need to know the maximal TCP window size to take the TCP congestion mechanism into account (see this page for details). On Linux, this value can be retrieved using the following commands. Both give a set of values, and you should use the last one, which is the maximal size.

$ cat /proc/sys/net/ipv4/tcp_rmem # gives the sender window
$ cat /proc/sys/net/ipv4/tcp_wmem # gives the receiver window

If you want to disable the TCP windowing mechanism, set this parameter to 0.

Manual calibration factors

SimGrid can take network irregularities such as a slow startup or changing behavior depending on the message size into account. The values provided by default were computed a long time ago through data fitting one the timings of either packet-level simulators or direct experiments on real platforms. These default values should be OK for most users, but if simulation realism is really important to you, you probably want to recalibrate the models (i.e., devise sensible values for your specific settings). This section only describes how to pass new values to the models while the calibration process involved in the computation of these values is described in the relevant chapter.

We found out that many networking effects can be realistically accounted for with the three following correction factors. They were shown to be enough to capture slow-start effects, the different transmission modes of MPI systems (eager vs. rendez-vous mode), or the non linear effects of wifi sharing.

Option network/latency-factor Default: 1.0, but overridden by most models

This option specifies a multiplier to apply to the physical latency (i.e., the one described in the platform) of the set of links involved in a communication. The factor can either be a constant to apply to any communication, or it may depend on the message size. The CM02 model does not use any correction factor, so the latency-factor remains to 1. The LV08 model sets it to 13.01 to model slow-start, while the SMPI model has several possible values depending on the interval in which the message size falls. The default SMPI setting given below specifies for example that a message smaller than 257 bytes will get a latency multiplier of 2.01467 while a message whose size is in [15424, 65472] will get a latency multiplier of 3.48845. The wifi model goes further and uses a callback in the program to compute the factor that must be non-linear in this case.

This multiplier is applied to the latency computed from the platform, that is the sum of all link physical latencies over the network path used by the considered communication, to derive the effective end-to-end latency.

Constant factors are easy to express, but the interval-based syntax used in SMPI is somewhat complex. It expects a set of factors separated by semicolons, each of the form boundary:factor. For example if your specification is 0:1;1000:2;5000:3, it means that on [0, 1000) the factor is 1. On [1000,5000), the factor is 2 while the factor is 3 for 5000 and beyond. If your first interval does include size=0, then the default value of 1 is used before. Changing the factor callback is not possible from the command line and must be done from your code, as shown in this example. Note that the chosen model only provides some default settings. You may pick a LV08 model to get some of the settings, and override the latency with interval-based values.

SMPI default value: 65472:11.6436; 15424:3.48845; 9376:2.59299; 5776:2.18796; 3484:1.88101; 1426:1.61075; 732:1.9503; 257:1.95341;0:2.01467 (interval boundaries are sorted automatically). These values were computed by data fitting on the Stampede Supercomputer at TACC, with optimal deployment of processes on nodes. To accurately model your settings, you should redo the calibration.

Option network/bandwidth-factor Default: 1.0, but overridden by most models

Setting this option automatically adjusts the effective bandwidth (i.e., the one perceived by the application) used by any given communication. As with latency-factor above, the value can be a constant (CM02 uses 1 – no correction – while LV08 uses 0.97 to discount TCP headers while computing the payload bandwidth), interval-based (as the default provided by the SMPI), or using in-program callbacks (as with wifi).

SMPI default value: 65472:0.940694;15424:0.697866;9376:0.58729;5776:1.08739;3484:0.77493;1426:0.608902;732:0.341987;257:0.338112;0:0.812084 This was also computed on the Stampede Supercomputer.

Option network/weight-S Default: depends on the model

Value used to account for RTT-unfairness when sharing a bottleneck (network connections with a large RTT are generally penalized against those with a small one). See The TCP models and also this scientific paper: Accuracy Study and Improvement of Network Simulation in the SimGrid Framework

Default values for CM02 is 0. LV08 sets it to 20537 while both SMPI and IB set it to 8775.

Infiniband model

InfiniBand network behavior can be modeled through 3 parameters smpi/IB-penalty-factors:"βe;βs;γs", as explained in the PhD thesis of Jérôme Vienne (in French) or more concisely in this paper, even if that paper does only describe models for myrinet and ethernet. You can see in Fig 2 some results for Infiniband, for example. This model may be outdated by now for modern infiniband, anyway, so a new validation would be good.

The three paramaters are defined as follows:

  • βs: penalty factor for outgoing messages, computed by running a simple send to two nodes and checking slowdown compared to a single send to one node, dividing by 2

  • βe: penalty factor for ingoing messages, same computation method but with one node receiving several messages

  • γr: slowdown factor when communication buffer memory is saturated. It needs a more complicated pattern to run in order to be computed (5.3 in the thesis, page 107), and formula in the end is γr = time(c)/(3×βe×time(ref)), where time(ref) is the time of a single comm with no contention).

Once these values are computed, a penalty is assessed for each message (this is the part implemented in the simulator) as shown page 106 of the thesis. Here is a simple translation of this text. First, some notations:

  • ∆e(e) which corresponds to the incoming degree of node e, that is to say the number of communications having as destination node e.

  • ∆s (s) which corresponds to the degree outgoing from node s, that is to say the number of communications sent by node s.

  • Φ (e) which corresponds to the number of communications destined for the node e but coming from a different node.

  • Ω (s, e) which corresponds to the number of messages coming from node s to node e. If node e only receives communications from different nodes then Φ (e) = ∆e (e). On the other hand if, for example, there are three messages coming from node s and going from node e then Φ (e) 6 = ∆e (e) and Ω (s, e) = 3

To determine the penalty for a communication, two values need to be calculated. First, the penalty caused by the conflict in transmission, noted ps.

  • if ∆s (i) = 1 then ps = 1.

  • if ∆s (i) ≥ 2 and ∆e (i) ≥ 3 then ps = ∆s (i) × βs × γr

  • else, ps = ∆s (i) × βs

Then, the penalty caused by the conflict in reception (noted pe) should be computed as follows:

  • if ∆e (i) = 1 then pe = 1

  • else, pe = Φ (e) × βe × Ω (s, e)

Finally, the penalty associated with the communication is: p = max (ps ∈ s, pe)

Simulating Cross-Traffic

Since SimGrid v3.7, cross-traffic effects can be taken into account in analytical simulations. It means that ongoing and incoming communication flows are treated independently. In addition, the LV08 model adds 0.05 of usage on the opposite direction for each new created flow. This can be useful to simulate some important TCP phenomena such as ack compression.

For that to work, your platform must have two links for each pair of interconnected hosts. An example of usable platform is available in examples/platforms/crosstraffic.xml.

This is activated through the network/crosstraffic item, that can be set to 0 (disable this feature) or 1 (enable it).

Note that with the default host model this option is activated by default.

Simulating Asynchronous Send

It is possible to specify that messages below a certain size (in bytes) will be sent as soon as the call to MPI_Send is issued, without waiting for the correspondent receive. This threshold can be configured through the smpi/async-small-thresh item. The default value is 0. This behavior can also be manually set for mailboxes, by setting the receiving mode of the mailbox with a call to sg_mailbox_set_receiver(). After this, all messages sent to this mailbox will have this behavior regardless of the message size.

This value needs to be smaller than or equals to the threshold set at Simulating MPI detached send, because asynchronous messages are meant to be detached as well.

Configuring ns-3

Option ns3/NetworkModel Default: “default” (ns-3 default TCP)

When using ns-3, the item ns3/NetworkModel can be used to switch between TCP or UDP, and switch the used TCP variante. If the item is left unchanged, ns-3 uses the default TCP implementation. With a value of “UDP”, ns-3 is set to use UDP instead. With the value of either ‘NewReno’ or ‘Cubic’, the ns3::TcpL4Protocol::SocketType configuration item in ns-3 is set to the corresponding protocol.

Option ns3/seed Default: “” (don’t set the seed in ns-3)

This option is the random seed to provide to ns-3 with ns3::RngSeedManager::SetSeed and ns3::RngSeedManager::SetRun.

If left blank, no seed is set in ns-3. If the value ‘time’ is provided, the current amount of seconds since epoch is used as a seed. Otherwise, the provided value must be a number to use as a seed.

Configuring the Storage model

File Descriptor Count per Host

Option storage/max_file_descriptors Default: 1024

Each host maintains a fixed-size array of its file descriptors. You can change its size through this item to either enlarge it if your application requires it or to reduce it to save memory space.

Activating Plugins

SimGrid plugins allow one to extend the framework without changing its source code directly. Read the source code of the existing plugins to learn how to do so (in src/plugins), and ask your questions to the usual channels (Stack Overflow, Mailing list, IRC). The basic idea is that plugins usually register callbacks to some signals of interest. If they need to store some information about a given object (Link, CPU or Actor), they do so through the use of a dedicated object extension.

Some of the existing plugins can be activated from the command line, meaning that you can activate them from the command line without any modification to your simulation code. For example, you can activate the host energy plugin by adding --cfg=plugin:host_energy to your command line.

Here is a partial list of plugins that can be activated this way. You can get the full list by passing --cfg=plugin:help to your simulator.

  • Host Energy: models the energy dissipation of the compute units.

  • Link Energy: models the energy dissipation of the network.

  • Host Load: monitors the load of the compute units.

Configuring the Model-Checking

To enable SimGrid’s model-checking support, the program should be executed using the simgrid-mc wrapper:

$ simgrid-mc ./my_program

Safety properties are expressed as assertions using the function void MC_assert(int prop)().

Specifying the MPI buffering behavior

Option smpi/buffering Default: infty

Buffering in MPI has a huge impact on the communication semantic. For example, standard blocking sends are synchronous calls when the system buffers are full while these calls can complete immediately without even requiring a matching receive call for small messages sent when the system buffers are empty.

In SMPI, this depends on the message size, that is compared against two thresholds:

  • if (size < smpi/async-small-thresh) then MPI_Send returns immediately, and the message is sent even if the corresponding receive has not be issued yet. This is known as the eager mode.

  • if (smpi/async-small-thresh < size < smpi/send-is-detached-thresh) then MPI_Send also returns immediately, but SMPI waits for the corresponding receive to be posted, in order to perform the communication operation.

  • if (smpi/send-is-detached-thresh < size) then MPI_Send returns only when the message has actually been sent over the network. This is known as the rendez-vous mode.

The smpi/buffering (only valid with MC) option gives an easier interface to choose between these semantics. It can take two values:

  • zero: means that buffering should be disabled. All communications are actually blocking.

  • infty: means that buffering should be made infinite. All communications are non-blocking.

Specifying the kind of reduction

Option model-check/reduction Default: “dpor”

The main issue when using the model-checking is the state space explosion. You can activate some reduction technique with --cfg=model-check/reduction:<technique>. For now, this configuration variable can take 2 values:

  • none: Do not apply any kind of reduction

  • dpor: Apply Dynamic Partial Ordering Reduction. Only valid if you verify local safety properties (default value for safety checks).

  • sdpor: Source-set DPOR, as described in “Source Sets: A Foundation for Optimal Dynamic Partial Order Reduction”

    by Abdulla et al.

  • odpor: Optimal DPOR, as described in “Source Sets: A Foundation for Optimal Dynamic Partial Order Reduction”

    by Abdulla et al.

Our current DPOR implementation could be improved in may ways. We are currently improving its efficiency (both in term of reduction ability and computational speed).

Guiding strategy

Option model-check/strategy Default: “none”

Even after the DPOR’s reduction, the state space that we have to explore remains huge. SimGrid provides several guiding strategies aiming at converging faster toward bugs. By default, none of these strategy is enabled, and SimGrid does a regular DFS exploration.

  • max_match_comm: Try to minimize the number of in-fly communication by appairing matching send and receive. This tend to produce nicer backtraces, in particular when a user-level send is broken down internally into a send_async + wait. This strategy will ensure that the wait occures as soon as possible, easing the understanding of the user who do not expect her send to be split.

  • min_match_comm: Try to maximize the number of in-fly communication by not appairing matching send and receive. This is the exact opposite strategy, but it is still useful as it tend to explore first the branches where the risk of deadlock is higher.

  • uniform: this is a boring random strategy where choices are based on a uniform sampling of possible choices. Surprisingly, it gives really really good results.

Dot Output

If set, the model-check/dot-output configuration item is the name of a file in which to write a dot file of the path leading to the property violation discovered (safety violation). This dot file can then be fed to the graphviz dot tool to generate a corresponding graphical representation.

Exploration Depth Limit

The model-check/max-depth can set the maximum depth of the exploration graph of the model checker. If this limit is reached, a logging message is sent and the results might not be exact.

By default, the exploration is limited to the depth of 1000.

Maximal amount of errors

The model-check/max-errors can be used to find more than one error in a given code. This may be useful if the trace of the first encountered faulty execution is too long. In that case, increasing the value of this option may help to find another trace that could be smaller. Using a negative value ensures exhaustive exploration, with no maximal amount on the number of found errors.

By default, the exploration stops after the first error (value = 0).

Searching for the critical transition

When it finds a failure, SimGrid automatically searches for the so-called critical transition. Before that transition, at least one exploration is correct; After it, all explorations are faulty. We hope that exhibiting critical transition will help you understanding the error. This option is enabled by default, unless when the model-checker is instructed to continue after the first error (in which case it cannot properly look for the critical transition).

Setting a timeout upon execution

If you wish to search for bugs in a best effort way, you probably want to explore with the uniform strategy (see cfg=model-check/strategy) and with a soft timeout. If such a timeout expires, the exploration will gracefully terminate at the next exploration end, without backtracking to a yet-to-be-explored case. The return code is 0 in this case.

Handling of Timeouts

By default, the model checker does not handle timeout conditions: the wait operations never time out. With the model-check/timeout configuration item set to yes, the model checker will explore timeouts of wait operations.

Communication Determinism

The model-check/communications-determinism and model-check/send-determinism items can be used to select the communication determinism mode of the model checker, which checks determinism properties of the communications of an application.

Exploration strategies

By default, the model checker follows a depth-first exploration, but several other strategies exist. The following list may be incomplete, so you’d better look at the code or speak with us for more information.

  • uniform: Randomly pick the next branch to explore each time that the exploration reaches an end.

  • none: default value instructing SimGrid to follow a depth-first exploration.

Caching states for performance

To explore new execution branches, the verified application must be rollback to its original state, and some transitions must be replayed to bring the application to the desired decision point. If the application induces many computations, replaying the transition from the beginning of the application can be time-consuming. To save time, one can use the cached-states-interval configuration item to save intermediate states that can be used as a starting point while restoring an applicative state. This is implemented by forking the executing application to later restart from that fork instead of from the begining.

By default, one in every 1000 states is cached this way. If the used value is 0, then no state gets cached. Caching too little states forces many useless transitions replays (consuming time) while caching too much states may exhaust the memory and other resources. Increasing the maximal amount of open file per process on your machine may allow to cache more states if your memory allows. States get removed from the memory once they become useless, so your resource consumption should plateau at some point during the exploration.

Verifying Python or multitheaded codes

By default, the model checker relies on system forks to speed-up the exploration but POSIX forbids the use of this system call in multithreaded applications. So, you cannot verify an application with Selecting the Virtualization Factory set to thread. Since our Python bindings unfortunately require the threaded contexts, this makes it impossible to use forks to speed up the verification of Python programs. In this case, use --cfg=model-check/no-fork:1 to go for the slow exploration without forks. With this option, a brand new python interpreter will be started when the model checker needs to rollback the application to explore another execution branch. This is slow, and will only work if your application is perfectly reproducible.

Passing environment variables

You can specify extra environment variables to be set in the verified application with model-check/setenv. For example, you can preload a library as follows: -cfg=model-check/setenv:LD_PRELOAD=toto;LD_LIBRARY_PATH=/tmp.

Verification Performance Considerations

The size of the stacks can have a huge impact on the memory consumption when using model-checking. By default, each snapshot will save a copy of the whole stacks and not only of the part that is really meaningful: you should expect the contribution of the memory consumption of the snapshots to be: \(\text{number of processes} \times \text{stack size} \times \text{number of states}\).

When compiled against the model checker, the stacks are not protected with guards: if the stack size is too small for your application, the stack will silently overflow into other parts of the memory (see contexts/guard-size).

Replaying buggy execution paths from the model checker

Debugging the problems reported by the model checker is challenging because the model checker may explore several execution paths before encountering the issue, the output very difficult to understand. Fortunately, SimGrid provides the execution path leading to any reported issue so that you can replay this path reported by the model checker, restoring a classical experience of debugging.

When the model checker finds an interesting path in the application execution graph (where a safety property is violated), it generates an identifier for this path. Here is an example of the output:

[  0.000000] (0:@) Check a safety property
[  0.000000] (0:@) **************************
[  0.000000] (0:@) *** PROPERTY NOT VALID ***
[  0.000000] (0:@) **************************
[  0.000000] (0:@) Counter-example execution trace:
[  0.000000] (0:@)   [(1)Tremblay (app)] MC_RANDOM(3)
[  0.000000] (0:@)   [(1)Tremblay (app)] MC_RANDOM(4)
[  0.000000] (0:@) Path = 1/3;1/4
[  0.000000] (0:@) Expanded states = 27
[  0.000000] (0:@) Visited states = 68
[  0.000000] (0:@) Executed transitions = 46

The interesting line is Path = 1/3;1/4, which means that you should use --cfg=model-check/replay:1/3;1/4 to replay your application on the buggy execution path. All options (but the model checker related ones) must remain the same. In particular, if you ran your application with smpirun -wrapper simgrid-mc, then remove the -wrapper simgrid-mc part (you may want to use valgrind or gdb as wrappers instead). Also remove all MC-related options, keep non-MC-related ones and add --cfg=model-check/replay:???.

Things are very similar if you are using sthread. Simply drop simgrid-mc from your command line, as follows:

$ LD_PRELOAD=../../lib/libsthread.so ./pthread-mutex-simpledeadlock --cfg=model-check/replay:'2;2;3;2;3;3'
sthread is intercepting the execution of ./pthread-mutex-simpledeadlock
[0.000000] [xbt_cfg/INFO] Configuration change: Set 'model-check/replay' to '2;2;3;2;3;3'
[0.000000] [mc_record/INFO] path=2;2;3;2;3;3
All threads are started.
[0.000000] [mc_record/INFO] ***********************************************************************************
[0.000000] [mc_record/INFO] * Path chunk #1 '2/0' Actor thread 1(pid:2): MUTEX_ASYNC_LOCK(mutex_id:0 owner:none)
[0.000000] [mc_record/INFO] ***********************************************************************************
Backtrace (displayed in actor thread 1):
  ->  #0 simgrid::s4u::Mutex::lock() at ../../src/s4u/s4u_Mutex.cpp:26
  ->  #1 sthread_mutex_lock at ../../src/sthread/sthread_impl.cpp:188
  ->  #2 pthread_mutex_lock at ../../src/sthread/sthread.c:141
  ->  #3 thread_fun1 at ../../examples/sthread/pthread-mutex-simpledeadlock.c:21

[0.000000] [mc_record/INFO] ***********************************************************************************
[0.000000] [mc_record/INFO] * Path chunk #2 '2/0' Actor thread 1(pid:2): MUTEX_WAIT(mutex_id:0 owner:2)
[0.000000] [mc_record/INFO] ***********************************************************************************
Backtrace (displayed in actor thread 1):
  ->  #0 simgrid::s4u::Mutex::lock() at ../../src/s4u/s4u_Mutex.cpp:29
  ->  #1 sthread_mutex_lock at ../../src/sthread/sthread_impl.cpp:188
  ->  #2 pthread_mutex_lock at ../../src/sthread/sthread.c:141
  ->  #3 thread_fun1 at ../../examples/sthread/pthread-mutex-simpledeadlock.c:21

[0.000000] [mc_record/INFO] ***********************************************************************************
[0.000000] [mc_record/INFO] * Path chunk #3 '3/0' Actor thread 2(pid:3): MUTEX_ASYNC_LOCK(mutex_id:1 owner:none)
[0.000000] [mc_record/INFO] ***********************************************************************************
Backtrace (displayed in actor thread 2):
  ->  #0 simgrid::s4u::Mutex::lock() at ../../src/s4u/s4u_Mutex.cpp:26
  ->  #1 sthread_mutex_lock at ../../src/sthread/sthread_impl.cpp:188
  ->  #2 pthread_mutex_lock at ../../src/sthread/sthread.c:141
  ->  #3 thread_fun2 at ../../examples/sthread/pthread-mutex-simpledeadlock.c:31

[0.000000] [mc_record/INFO] ***********************************************************************************
[0.000000] [mc_record/INFO] * Path chunk #4 '2/0' Actor thread 1(pid:2): MUTEX_ASYNC_LOCK(mutex_id:1 owner:3)
[0.000000] [mc_record/INFO] ***********************************************************************************
Backtrace (displayed in actor thread 1):
  ->  #0 simgrid::s4u::Mutex::lock() at ../../src/s4u/s4u_Mutex.cpp:26
  ->  #1 sthread_mutex_lock at ../../src/sthread/sthread_impl.cpp:188
  ->  #2 pthread_mutex_lock at ../../src/sthread/sthread.c:141
  ->  #3 thread_fun1 at ../../examples/sthread/pthread-mutex-simpledeadlock.c:22

[0.000000] [mc_record/INFO] ***********************************************************************************
[0.000000] [mc_record/INFO] * Path chunk #5 '3/0' Actor thread 2(pid:3): MUTEX_WAIT(mutex_id:1 owner:3)
[0.000000] [mc_record/INFO] ***********************************************************************************
Backtrace (displayed in actor thread 2):
  ->  #0 simgrid::s4u::Mutex::lock() at ../../src/s4u/s4u_Mutex.cpp:29
  ->  #1 sthread_mutex_lock at ../../src/sthread/sthread_impl.cpp:188
  ->  #2 pthread_mutex_lock at ../../src/sthread/sthread.c:141
  ->  #3 thread_fun2 at ../../examples/sthread/pthread-mutex-simpledeadlock.c:31

[0.000000] [mc_record/INFO] ***********************************************************************************
[0.000000] [mc_record/INFO] * Path chunk #6 '3/0' Actor thread 2(pid:3): MUTEX_ASYNC_LOCK(mutex_id:0 owner:2)
[0.000000] [mc_record/INFO] ***********************************************************************************
Backtrace (displayed in actor thread 2):
  ->  #0 simgrid::s4u::Mutex::lock() at ../../src/s4u/s4u_Mutex.cpp:26
  ->  #1 sthread_mutex_lock at ../../src/sthread/sthread_impl.cpp:188
  ->  #2 pthread_mutex_lock at ../../src/sthread/sthread.c:141
  ->  #3 thread_fun2 at ../../examples/sthread/pthread-mutex-simpledeadlock.c:32

[0.000000] [mc_record/INFO] The replay of the trace is complete. DEADLOCK detected.
[0.000000] [ker_engine/INFO] 3 actors are still running, waiting for something.
[0.000000] [ker_engine/INFO] Legend of the following listing: "Actor <pid> (<name>@<host>): <status>"
[0.000000] [ker_engine/INFO] Actor 1 (main thread@Lilibeth) simcall ActorJoin(pid:2)
[0.000000] [ker_engine/INFO] Actor 2 (thread 1@Lilibeth) simcall MUTEX_WAIT(mutex_id:1 owner:3)
[0.000000] [ker_engine/INFO] Actor 3 (thread 2@Lilibeth) simcall MUTEX_WAIT(mutex_id:0 owner:2)
[0.000000] [sthread/INFO] All threads exited. Terminating the simulation.
[0.000000] ../../src/kernel/EngineImpl.cpp:265: [ker_engine/WARNING] Process called exit when leaving - Skipping cleanups
[0.000000] ../../src/kernel/EngineImpl.cpp:265: [ker_engine/WARNING] Process called exit when leaving - Skipping cleanups

Currently, if the path is of the form X;Y;Z, each number denotes the actor’s pid that is selected at each indecision point. If it’s of the form X/a;Y/b, the X and Y are the selected pids while the a and b are the return values of their simcalls. In the previous example, 1/3;1/4, you can see from the full output that the actor 1 is doing MC_RANDOM simcalls, so the 3 and 4 simply denote the values that these simcall return on the execution branch leading to the violation.

Configuring the User Code Virtualization

Selecting the Virtualization Factory

Option contexts/factory Default: “raw”

In SimGrid, the user code is virtualized in a specific mechanism that allows the simulation kernel to control its execution: when a user process requires a blocking action (such as sending a message), it is interrupted, and only gets released when the simulated clock reaches the point where the blocking operation is done. This is explained graphically in the relevant tutorial, available online.

In SimGrid, the containers in which user processes are virtualized are called contexts. Several context factory are provided, and you can select the one you want to use with the contexts/factory configuration item. Some of the following may not exist on your machine because of portability issues. In any case, the default one should be the most effcient one (please report bugs if the auto-detection fails for you). They are approximately sorted here from the slowest to the most efficient:

  • thread: very slow factory using full featured, standard threads. They are slow but very standard. Some debuggers or profilers only work with this factory.

  • ucontext: fast factory using System V contexts (Linux and FreeBSD only)

  • boost: This uses the context implementation of the boost library for a performance that is comparable to our raw implementation.
    Install the relevant library (e.g. with the libboost-contexts-dev package on Debian/Ubuntu) and recompile SimGrid.

  • raw: amazingly fast factory using a context switching mechanism of our own, directly implemented in assembly (only available for x86 and amd64 platforms for now) and without any unneeded system call.

The main reason to change this setting is when the debugging tools become fooled by the optimized context factories. Threads are the most debugging-friendly contexts, as they allow one to set breakpoints anywhere with gdb and visualize backtraces for all processes, in order to debug concurrency issues. Valgrind is also more comfortable with threads, but it should be usable with all factories (Exception: the callgrind tool really dislikes raw and ucontext factories).

Adapting the Stack Size

Option contexts/stack-size Default: 8192 KiB

Each virtualized used process is executed using a specific system stack. The size of this stack has a huge impact on the simulation scalability, but its default value is rather large. This is because the error messages that you get when the stack size is too small are rather disturbing: this leads to stack overflow (overwriting other stacks), leading to segfaults with corrupted stack traces.

If you want to push the scalability limits of your code, you might want to reduce the contexts/stack-size item. Its default value is 8192 (in KiB), while our Chord simulation works with stacks as small as 16 KiB, for example. You can ensure that some actors have a specific size by simply changing the value of this configuration item before creating these actors. The simgrid::s4u::Engine::set_config() functions are handy for that.

This setting is ignored when using the thread factory (because there is no way to modify the stack size with C++ system threads). Instead, you should compile SimGrid and your application with -fsplit-stack. Note that this compilation flag is not compatible with the model checker right now.

The operating system should only allocate memory for the pages of the stack which are actually used and you might not need to use this in most cases. However, this setting is very important when using the model checker (see Verification Performance Considerations).

Disabling Stack Guard Pages

Option contexts/guard-size Default 1 page in most case (0 pages with MC)

Unless you use the threads context factory (see Selecting the Virtualization Factory), a stack guard page is usually used which prevents the stack of a given actor from overflowing on another stack. But the performance impact may become prohibitive when the amount of actors increases. The option contexts/guard-size is the number of stack guard pages used. By setting it to 0, no guard pages will be used: in this case, you should avoid using small stacks (with contexts/stack-size) as the stack will silently overflow on other parts of the memory.

When no stack guard page is created, stacks may then silently overflow on other parts of the memory if their size is too small for the application.

Running User Code in Parallel

Parallel execution of the user code is only considered stable in SimGrid v3.7 and higher, and mostly for S4U simulations. SMPI simulations may well fail in parallel mode. It is described in INRIA RR-7653.

Note that this feature is only tested on Linux. It may or may not work on other systems.

If you are using the ucontext or raw context factories, you can request to execute the user code in parallel. Several threads are launched, each of them handling the same number of user contexts at each run. To activate this, set the contexts/nthreads item to the amount of cores that you have in your computer (or lower than 1 to have the amount of cores auto-detected).

When parallel execution is activated, you can choose the synchronization schema used with the contexts/synchro item, which value is either:

  • futex: ultra optimized synchronisation schema, based on futexes (fast user-mode mutexes), and thus only available on Linux systems. This is the default mode when available.

  • posix: slow but portable synchronisation using only POSIX primitives.

  • busy_wait: not really a synchronisation: the worker threads constantly request new contexts to execute. It should be the most efficient synchronisation schema, but it loads all the cores of your machine for no good reason. You probably prefer the other less eager schemas.

Configuring the Tracing

The tracing subsystem can be configured in several different ways depending on the used interface (S4U, SMPI) and the kind of traces that needs to be obtained. See the Tracing Configuration Options subsection for a full description of each configuration option.

We detail here a simple way to get the traces working for you, even if you never used the tracing API.

  • Any SimGrid-based simulator (S4U, SMPI, …) and raw traces:

    --cfg=tracing:yes --cfg=tracing/uncategorized:yes
    

    The first parameter activates the tracing subsystem, and the second tells it to trace host and link utilization (without any categorization).

  • S4U-based simulator and categorized traces (you need to declare categories and classify your tasks according to them)

    --cfg=tracing:yes --cfg=tracing/categorized:yes
    

    The first parameter activates the tracing subsystem, and the second tells it to trace host and link categorized utilization.

  • SMPI simulator and traces for a space/time view:

    $ smpirun -trace ...
    

    The -trace parameter for the smpirun script runs the simulation with --cfg=tracing:yes --cfg=tracing/smpi:yes. Check the smpirun’s -help parameter for additional tracing options.

Sometimes you might want to put additional information on the trace to correctly identify them later, or to provide data that can be used to reproduce an experiment. You have two ways to do that:

  • Add a string on top of the trace file as comment:

    --cfg=tracing/comment:my_simulation_identifier
    
  • Add the contents of a textual file on top of the trace file as comment:

    --cfg=tracing/comment-file:my_file_with_additional_information.txt
    

Please, use these two parameters (for comments) to make reproducible simulations. For additional details about this and all tracing options, check See the tracing_tracing_options.

Configuring SMPI

The SMPI interface provides several specific configuration items. These are not easy to see with --help-cfg, since SMPI binaries are usually launched through the smiprun script.

Automatic Benchmarking of SMPI Code

In SMPI, the sequential code is automatically benchmarked, and these computations are automatically reported to the simulator. That is to say that if you have a large computation between a MPI_Recv() and a MPI_Send(), SMPI will automatically benchmark the duration of this code, and create an execution task within the simulator to take this into account. For that, the actual duration is measured on the host machine and then scaled to the power of the corresponding simulated machine. The variable smpi/host-speed allows one to specify the computational speed of the host machine (in flop/s by default) to use when scaling the execution times.

The default value is smpi/host-speed=20kf (= 20,000 flop/s). This is probably underestimated for most machines, leading SimGrid to overestimate the amount of flops in the execution blocks that are automatically injected in the simulator. As a result, the execution time of the whole application will probably be overestimated until you use a realistic value.

When the code consists of numerous consecutive MPI calls, the previous mechanism feeds the simulation kernel with numerous tiny computations. The smpi/cpu-threshold item becomes handy when this impacts badly on the simulation performance. It specifies a threshold (in seconds) below which the execution chunks are not reported to the simulation kernel (default value: 1e-6).

Note

The option smpi/cpu-threshold ignores any computation time spent below this threshold. SMPI does not consider the amount of time of these computations; there is no offset for this. Hence, a value that is too small, may lead to unreliable simulation results.

In some cases, however, one may wish to disable simulation of the computation of an application. This is the case when SMPI is used not to simulate an MPI application, but instead an MPI code that performs “live replay” of another MPI app (e.g., ScalaTrace’s replay tool, or various on-line simulators that run an app at scale). In this case the computation of the replay/simulation logic should not be simulated by SMPI. Instead, the replay tool or on-line simulator will issue “computation events”, which correspond to the actual MPI simulation being replayed/simulated. At the moment, these computation events can be simulated using SMPI by calling internal smpi_execute*() functions.

To disable the benchmarking/simulation of a computation in the simulated application, the variable smpi/simulate-computation should be set to no. This option just ignores the timings in your simulation; it still executes the computations itself. If you want to stop SMPI from doing that, you should check the SMPI_SAMPLE macros, documented in Section Toward Faster Simulations.

Solution

Computations executed?

Computations simulated?

–cfg=smpi/simulate-computation:no

Yes

Never

–cfg=smpi/cpu-threshold:42

Yes, in all cases

If it lasts over 42 seconds

SMPI_SAMPLE() macro

Only once per loop nest

Always

Slow-down or speed-up parts of your code

Option smpi/comp-adjustment-file: Default: unset

This option allows you to pass a file that contains two columns: The first column defines the section that will be subject to a speedup; the second column is the speedup. For instance:

"start:stop","ratio"
"exchange_1.f:30:exchange_1.f:130",1.18244559422142

The first line is the header - you must include it. The following line means that the code between two consecutive MPI calls on line 30 in exchange_1.f and line 130 in exchange_1.f should receive a speedup of 1.18244559422142. The value for the second column is therefore a speedup, if it is larger than 1 and a slowdown if it is smaller than 1. Nothing will be changed if it is equal to 1.

Of course, you can set any arbitrary filenames you want (so the start and end don’t have to be in the same file), but be aware that this mechanism only supports consecutive calls!

Please note that you must pass the -trace-call-location flag to smpicc or smpiff, respectively. This flag activates some internal macro definitions that help with obtaining the call location.

Bandwidth and latency factors

Adapting the bandwidth and latency acurately to the network conditions is of a paramount importance to get realistic results. This is done through the network/bandwidth-factor and network/latency-factor items. You probably also want to read the following section: Calibrating the models.

Reporting Simulation Time

Option smpi/display-timing Default: 0 (false)

Most of the time, you run MPI code with SMPI to compute the time it would take to run it on a platform. But since the code is run through the smpirun script, you don’t have any control on the launcher code, making it difficult to report the simulated time when the simulation ends. If you enable the smpi/display-timing item, smpirun will display this information when the simulation ends. SMPI will also display information about the amout of real time spent in application code and in SMPI internals, to provide hints about the need to use sampling to reduce simulation time.

Reporting memory allocations

Option smpi/display-allocs Default: 0 (false)

SMPI intercepts malloc and calloc calls performed inside the running application, if it wasn’t compiled with SMPI_NO_OVERRIDE_MALLOC. With this option, SMPI will show at the end of execution the amount of memory allocated through these calls, and locate the most expensive one. This helps finding the targets for manual memory sharing, or the threshold to use for smpi/auto-shared-malloc-thresh option (see Automatically share allocations).

Keeping temporary files after simulation

Option smpi/keep-temps default: 0 (false)

SMPI usually generates a lot of temporary files that are cleaned after use. This option requests to preserve them, for example to debug or profile your code. Indeed, the binary files are removed very early under the dlopen privatization schema, which tends to fool the debuggers.

Trace hardware counters with PAPI

Option smpi/papi-events default: unset

When the PAPI support is compiled into SimGrid, this option takes the names of PAPI counters and adds their respective values to the trace files (See Section tracing_tracing_options).

Warning

This feature currently requires superuser privileges, as registers are queried. Only use this feature with code you trust! Call smpirun for instance via smpirun -wrapper "sudo " <your-parameters> or run sudo sh -c "echo 0 > /proc/sys/kernel/perf_event_paranoid" In the later case, sudo will not be required.

It is planned to make this feature available on a per-process (or per-thread?) basis. The first draft, however, just implements a “global” (i.e., for all processes) set of counters, the “default” set.

--cfg=smpi/papi-events:"default:PAPI_L3_LDM:PAPI_L2_LDM"

Automatic Privatization of Global Variables

Option smpi/privatization default: “dlopen” (when using smpirun)

MPI executables are usually meant to be executed in separate processes, but SMPI is executed in only one process. Global variables from executables will be placed in the same memory region and shared between processes, causing intricate bugs. Several options are possible to avoid this, as described in the main SMPI publication and in the SMPI documentation. SimGrid provides two ways of automatically privatizing the globals, and this option allows one to choose between them.

  • no (default when not using smpirun): Do not automatically privatize variables. Pass -no-privatize to smpirun to disable this feature.

  • dlopen or yes (default when using smpirun): Link multiple times against the binary.

  • mmap (slower, but maybe somewhat more stable): Runtime automatic switching of the data segments.

Warning

This configuration option cannot be set in your platform file. You can only pass it as an argument to smpirun.

Automatic privatization of global variables inside external libraries

Option smpi/privatize-libs default: unset

Linux/BSD only: When using dlopen (default) privatization, privatize specific shared libraries with internal global variables, if they can’t be linked statically. For example libgfortran is usually used for Fortran I/O and indexes in files can be mixed up.

Multiple libraries can be given, semicolon separated.

This configuration option can only use either full paths to libraries, or full names. Check with ldd the name of the library you want to use. For example:

$ ldd allpairf90
   ...
   libgfortran.so.3 => /usr/lib/x86_64-linux-gnu/libgfortran.so.3 (0x00007fbb4d91b000)
   ...

Then you can use --cfg=smpi/privatize-libs:libgfortran.so.3 or --cfg=smpi/privatize-libs:/usr/lib/x86_64-linux-gnu/libgfortran.so.3, but not libgfortran nor libgfortran.so.

Simulating MPI detached send

Option smpi/send-is-detached-thresh default: 65536

This threshold specifies the size in bytes under which the send will return immediately. This is different from the threshold detailed in Simulating Asynchronous Send because the message is not really sent when the send is posted. SMPI still waits for the corresponding receive to be posted, in order to perform the communication operation.

Simulating MPI collective algorithms

Option smpi/coll-selector Possible values: naive (default), ompi, mpich

SMPI implements more than 100 different algorithms for MPI collective communication, to accurately simulate the behavior of most of the existing MPI libraries. The smpi/coll-selector item can be used to select the decision logic either of the OpenMPI or the MPICH libraries. (By default SMPI uses naive version of collective operations.)

Each collective operation can be manually selected with a smpi/collective_name:algo_name. For example, if you want to use the Bruck algorithm for the Alltoall algorithm, you should use --cfg=smpi/alltoall:bruck on the command-line of smpirun. The reference of all available algorithms are listed in Simulating Collective Operations, and you can get the full list implemented in your version using smpirun --help-coll.

Add a barrier in all collectives

Option smpi/barrier-collectives default: off

This option adds a simple barrier in some collective operations to catch dangerous code that may or may not work depending on the MPI implementation: Bcast, Exscan, Gather, Gatherv, Scan, Scatter, Scatterv and Reduce.

For example, the following code works with OpenMPI while it deadlocks in MPICH and Intel MPI. Broadcast seem to be “fire and forget” in OpenMPI while other implementations expect to receive a message.

if (rank == 0) {
  MPI_Bcast(buf1, buff_size, MPI_CHAR, 0, newcom);
  MPI_Send(&buf2, buff_size, MPI_CHAR, 1, tag, newcom);
} else if (rank==1) {
  MPI_Recv(&buf2, buff_size, MPI_CHAR, 0, tag, newcom, MPI_STATUS_IGNORE);
  MPI_Bcast(buf1, buff_size, MPI_CHAR, 0, newcom);
}

The barrier is only simulated and does not involve any additional message (it is a S4U barrier). This option is disabled by default, and activated by the -analyze flag of smpirun.

Add a barrier in MPI_Finalize

Option smpi/finalization-barrier default: off

By default, SMPI processes are destroyed as soon as soon as their code ends, so after a successful MPI_Finalize call returns. In some rare cases, some data might have been attached to MPI objects still active in the remaining processes, and can be destroyed eagerly by the finished process. If your code shows issues at finalization, such as segmentation fault, triggering this option will add an explicit MPI_Barrier(MPI_COMM_WORLD) call inside the MPI_Finalize, so that all processes will terminate at almost the same point. It might affect the total timing by the cost of a barrier.

Disable MPI fatal errors

Option smpi/errors-are-fatal default: on

By default, SMPI processes will crash if a MPI error code is returned. MPI allows to explicitely set MPI_ERRORS_RETURN errhandler to avoid this behaviour. This flag will turn on this behaviour by default (for all concerned types and errhandlers). This can ease debugging by going after the first reported error.

Disable pedantic MPI errors

Option smpi/pedantic default: on

By default, SMPI will report all errors it finds in MPI codes. Some of these errors may not be considered as errors by all developers. This flag can be turned off to avoid reporting some usually harmless mistakes. Concerned errors list (will be expanded in the future):

  • Calling MPI_Win_fence only once in a program, hence just opening an epoch without ever closing it.

Inject constant times for MPI_Iprobe

Option smpi/iprobe default: 0.0001

The behavior and motivation for this configuration option is identical with smpi/test, but for the function MPI_Iprobe()

Reduce speed for iprobe calls

Option smpi/iprobe-cpu-usage default: 1 (no change)

MPI_Iprobe calls can be heavily used in applications. To account correctly for the energy that cores spend probing, it is necessary to reduce the load that these calls cause inside SimGrid.

For instance, we measured a maximum power consumption of 220 W for a particular application but only 180 W while this application was probing. Hence, the correct factor that should be passed to this option would be 180/220 = 0.81.

Inject constant times for MPI_Init

Option smpi/init default: 0

The behavior and motivation for this configuration option is identical with smpi/test, but for the function MPI_Init().

Inject constant times for MPI_Isend()

Option smpi/ois

The behavior and motivation for this configuration option is identical with smpi/os, but for the function MPI_Isend().

Inject constant times for MPI_send()

Option smpi/os

In several network models such as LogP, send (MPI_Send, MPI_Isend) and receive (MPI_Recv) operations incur costs (i.e., they consume CPU time). SMPI can factor these costs in as well, but the user has to configure SMPI accordingly as these values may vary by machine. This can be done by using smpi/os for MPI_Send operations; for MPI_Isend and MPI_Recv, use smpi/ois and smpi/or, respectively. These work exactly as smpi/ois.

This item can consist of multiple sections; each section takes three values, for example 1:3:2;10:5:1. The sections are divided by “;” so this example contains two sections. Furthermore, each section consists of three values.

  1. The first value denotes the minimum size in bytes for this section to take effect; read it as “if message size is greater than this value (and other section has a larger first value that is also smaller than the message size), use this”. In the first section above, this value is “1”.

  2. The second value is the startup time; this is a constant value that will always be charged, no matter what the size of the message. In the first section above, this value is “3”.

  3. The third value is the per-byte cost. That is, it is charged for every byte of the message (incurring cost messageSize*cost_per_byte) and hence accounts also for larger messages. In the first section of the example above, this value is “2”.

Now, SMPI always checks which section it should use for a given message; that is, if a message of size 11 is sent with the configuration of the example above, only the second section will be used, not the first, as the first value of the second section is closer to the message size. Hence, when smpi/os=1:3:2;10:5:1, a message of size 11 incurs the following cost inside MPI_Send: 5+11*1 because 5 is the startup cost and 1 is the cost per byte.

Note that the order of sections can be arbitrary; they will be ordered internally.

Inject constant times for MPI_Recv()

Option smpi/or

The behavior and motivation for this configuration option is identical with smpi/os, but for the function MPI_Recv().

Inject constant times for MPI_Test

Option smpi/test default: 0.0001

By setting this option, you can control the amount of time a process sleeps when MPI_Test() is called; this is important, because SimGrid normally only advances the time while communication is happening and thus, MPI_Test will not add to the time, resulting in deadlock if it is used as a break-condition as in the following example:

while(!flag) {
    MPI_Test(request, flag, status);
    ...
}

To speed up execution, we use a counter to keep track of how often we checked if the handle is now valid or not. Hence, we actually use counter*SLEEP_TIME, that is, the time MPI_Test() causes the process to sleep increases linearly with the number of previously failed tests. This behavior can be disabled by setting smpi/grow-injected-times to no. This will also disable this behavior for MPI_Iprobe.

Factorize malloc()s

Option smpi/shared-malloc Possible values: global (default), local

If your simulation consumes too much memory, you may want to modify your code so that the working areas are shared by all MPI ranks. For example, in a block-cyclic matrix multiplication, you will only allocate one set of blocks, and all processes will share them. Naturally, this will lead to very wrong results, but this will save a lot of memory. So this is still desirable for some studies. For more on the motivation for that feature, please refer to the relevant section of the SMPI CourseWare (see Activity #2.2 of the pointed assignment). In practice, change the calls for malloc() and free() into SMPI_SHARED_MALLOC() and SMPI_SHARED_FREE().

SMPI provides two algorithms for this feature. The first one, called local, allocates one block per call to SMPI_SHARED_MALLOC() (each call site gets its own block) ,and this block is shared among all MPI ranks. This is implemented with the shm_* functions to create a new POSIX shared memory object (kept in RAM, in /dev/shm) for each shared block.

With the global algorithm, each call to SMPI_SHARED_MALLOC() returns a new address, but it only points to a shadow block: its memory area is mapped on a 1 MiB file on disk. If the returned block is of size N MiB, then the same file is mapped N times to cover the whole block. At the end, no matter how many times you call SMPI_SHARED_MALLOC, this will only consume 1 MiB in memory.

You can disable this behavior and come back to regular mallocs (for example for debugging purposes) using no as a value.

If you want to keep private some parts of the buffer, for instance if these parts are used by the application logic and should not be corrupted, you can use SMPI_PARTIAL_SHARED_MALLOC(size, offsets, offsets_count). For example:

mem = SMPI_PARTIAL_SHARED_MALLOC(500, {27,42 , 100,200}, 2);

This will allocate 500 bytes to mem, such that mem[27..41] and mem[100..199] are shared while other area remain private.

Then, it can be deallocated by calling SMPI_SHARED_FREE(mem).

When smpi/shared-malloc:global is used, the memory consumption problem is solved, but it may induce too much load on the kernel’s pages table. In this case, you should use huge pages so that the kernel creates only one entry per MB of malloced data instead of one entry per 4 kB. To activate this, you must mount a hugetlbfs on your system and allocate at least one huge page:

$ mkdir /home/huge
$ sudo mount none /home/huge -t hugetlbfs -o rw,mode=0777
$ sudo sh -c 'echo 1 > /proc/sys/vm/nr_hugepages' # echo more if you need more

Then, you can pass the option --cfg=smpi/shared-malloc-hugepage:/home/huge to smpirun to actually activate the huge page support in shared mallocs.

Automatically share allocations

Option smpi/auto-shared-malloc-thresh: Default: 0 (false)

This value in bytes represents the size above which all allocations will be “shared” by default (as if they were performed through SMPI_SHARED_MALLOC macros). Default = 0 = disabled feature. The value must be carefully chosen to only select data buffers which will not modify execution path or cause crash if their content is false. Option Reporting memory allocations can be used to locate the largest allocation detected in a run, and provide a good starting threshold. Note : malloc, calloc and free are overridden by smpicc/cxx by default. This can cause some troubles if codes are already overriding these. If this is the case, defining SMPI_NO_OVERRIDE_MALLOC in the compilation flags can help, but will make this feature unusable.

Inject constant times for MPI_Wtime, gettimeofday and clock_gettime

Option smpi/wtime default: 10 ns

This option controls the amount of (simulated) time spent in calls to MPI_Wtime(), gettimeofday() and clock_gettime(). If you set this value to 0, the simulated clock is not advanced in these calls, which leads to issues if your application contains such a loop:

while(MPI_Wtime() < some_time_bound) {
     /* some tests, with no communication nor computation */
}

When the option smpi/wtime is set to 0, the time advances only on communications and computations. So the previous code results in an infinite loop: the current [simulated] time will never reach some_time_bound. This infinite loop is avoided when that option is set to a small value, as it is by default since SimGrid v3.21.

Note that if your application does not contain any loop depending on the current time only, then setting this option to a non-zero value will slow down your simulations by a tiny bit: the simulation loop has to be broken out of and reset each time your code asks for the current time. If the simulation speed really matters to you, you can avoid this extra delay by setting smpi/wtime to 0.

Report leaked MPI objects

Option smpi/list-leaks default: 0

This option controls whether to report leaked MPI objects. The parameter is the number of leaks to report.

Other Configurations

Cleanup at Termination

Option debug/clean-atexit default: on

If your code is segfaulting during its finalization, it may help to disable this option to request that SimGrid not attempt any cleanups at the end of the simulation. Since the Unix process is ending anyway, the operating system will wipe it all.

Search Path

Option path default: . (current dir)

It is possible to specify a list of directories to search in for the trace files (see pf_trace) by using this configuration item. To add several directory to the path, set the configuration item several times, as in --cfg=path:toto --cfg=path:tutu

Set a Breakpoint

Option debug/breakpoint default: unset

This configuration option sets a breakpoint: when the simulated clock reaches the given time, a SIGTRAP is raised. This can be used to stop the execution and get a backtrace with a debugger.

It is also possible to set the breakpoint from inside the debugger, by writing in global variable simgrid::kernel::cfg_breakpoint. For example, with gdb:

set variable simgrid::kernel::cfg_breakpoint = 3.1416

Behavior on Ctrl-C

Option debug/verbose-exit default: on

By default, when Ctrl-C is pressed, the status of all existing actors is displayed before exiting the simulation. This is very useful to debug your code, but it can become troublesome if you have many actors. Set this configuration item to off to disable this feature.

Truncate local path from exception backtrace

Option exception/cutpath default: off

This configuration option is used to remove the path from the backtrace shown when an exception is thrown. This is mainly useful for the tests: the full file path would makes the tests non-reproducible because the paths of source files depend of the build settings. That would break most of the tests since their output is continually compared.

Logging configuration

As introduced in Textual logging, the SimGrid logging mechanism allows to configure at runtime the messages that should be displayed and those that should be omitted. Each message produced in the code is given a category (denoting its topic) and a priority. Then at runtime, each category is given a threshold (only messages of priority higher than that threshold are displayed), a layout (deciding how the messages in this category are formatted), and an appender (deciding what to do with the message: either print on stderr or to a file).

This section explains how to configure this logging features. You can also refer to the documentation of the programmer’s interface, that allows to produce messages from your code.

Most of the time, the logging mechanism is configured at runtime using the --log command-line argument, even if you can also use xbt_log_control_set() to control it from your program. To pass configure more than one setting, you can either pass several --log arguments, or separate your settings with spaces, that must be quoted accordingly. In practice, the following is equivalent to the above settings: --log=root.thresh:error --log=s4u_host.thresh:debug.

If you want to specify more than one setting, you can either pass several --log argument to your program as above, or separate them with spaces. In this case, you want to quote your settings, as in --log="root.thresh:error s4u_host.thresh:debug". The parameters are interpreted in order, from left to right.

Threshold configuration

The keyword threshold controls which logging event will get displayed in a given category. For example, --log=root.threshold:debug displays every message produced in the root category and its subcategories (i.e., every message produced – this is extremely verbose), while --log=root.thres:critical turns almost everything off. As you can see, threshold can be abbreviated here.

Existing thresholds:

  • trace some functions display a message at this level when entering or returning

  • debug output that is mostly useful when debugging the corresponding module.

  • verbose verbose output that is only mildly interesting and can easily be ignored

  • info usual output (this is the default threshold of all categories)

  • warning minor issue encountered

  • error issue encountered

  • critical major issue encountered, such as assertions failures

Format configuration

The keyword fmt controls the layout (the format) of a logging category. For example, --log=root.fmt:%m reduces the output to the user-message only, removing any decoration such as the date, or the actor ID, everything. Existing format directives:

  • %%: the % char

  • %n: line separator (LOG4J compatible)

  • %e: plain old space (SimGrid extension)

  • %m: user-provided message

  • %c: Category name (LOG4J compatible)

  • %p: Priority name (LOG4J compatible)

  • %h: Hostname (SimGrid extension)

  • %a: Actor name (SimGrid extension – note that with SMPI this is the integer value of the process rank)

  • %i: Actor PID (SimGrid extension – this is a ‘i’ as in ‘i’dea)

  • %I: system PID (the result of the UNIX getpid() function – only useful if your simulation forks at the system level)

  • %t: Thread “name” (LOG4J compatible – actually the address of the thread in memory)

  • %F: file name where the log event was raised (LOG4J compatible)

  • %l: location where the log event was raised (LOG4J compatible, like ‘%%F:%%L’ – this is a l as in ‘l’etter)

  • %L: line number where the log event was raised (LOG4J compatible)

  • %M: function name (LOG4J compatible – called method name here of course).

  • %d: date (UNIX-like epoch)

  • %r: application age (time elapsed since the beginning of the application)

--log=root.fmt:'[%h:%a:(%i) %r] %l: %m%n' gives you the default layout used for info messages while --log=root.fmt:'[%h:%a:(%i) %r] %l: [%c/%p] %m%n' gives you the default layout for the other priorities (it adds the source code location). Also, the actor identification is omitted by the default layout for the messages coming directly from the SimGrid kernel, so info messages are formatted with [%r] [%c/%p] %m%n in this case. When specifying the layout manually, such distinctions are currently impossible, and the provided layout is used for every messages.

As with printf, you can specify the precision and width of the fields. For example, %.4r limits the date precision to four digits while %15h limits the host name to at most 15 chars.

If you want to have spaces in your log format, you should protect it. Otherwise, SimGrid will consider that this is a space-separated list of several parameters. But you should also protect it from the shell that also splits command line arguments on spaces. At the end, you should use something such as --log="'root.fmt:%l: [%p/%c]: %m%n'". Another option is to use the %e directive for spaces, as in --log=root.fmt:%l:%e[%p/%c]:%e%m%n.

Category appender

The keyword app controls the appended of a logging category. For example --log=root.app:file:mylogfile redirects every output to the file mylogfile.

With the splitfile appender, a new file is created when the size of the output reaches the specified size. The format is --log=root.app:splitfile:<size>:<file name>. For example, --log=root.app:splitfile:500:mylog_% creates log files of at most 500 bytes, using the names mylog_0, mylog_1, mylog_2, etc.

The rollfile appender uses one file only, but the file is emptied and recreated when its size reaches the specified maximum. For example, --log=root.app:rollfile:500:mylog ensures that the log file mylog will never overpass 500 bytes in size.

Any appender setup this way have its own layout format, that you may change afterward. When specifying a new appender, its additivity is set to false to prevent log event displayed by this appender to “leak” to any other appender higher in the hierarchy. You can naturally change that if you want your messages to be displayed twice.

Category additivity

The keyword add controls the additivity of a logging category. By default, the messages are only passed one appender only: the more specific, i.e. the first one found when climbing the tree from the category in which they were produced. In Log4J parlance, it is said that the default additivity of appenders is false. If you change this setting to on (or yes or 1), the produced messages will also be passed to the upper appender.

Let’s consider a more complex example: --log="root.app:file:all.log s4u.app:file:iface.log xbt.app:file:xbt.log xbt.add:yes. Here, the logging of s4u will be sent to the iface.log file; the logging of the xbt toolbox will be sent to both the xbt.log file and the all.log file (because xbt additivity was enabled); and every other loggings will only be sent to all.log.

Other options

--help-logs displays a complete help message about logging in SimGrid.

--help-log-categories displays the actual hierarchy of log categories for this binary.

--log=no_loc hides the source locations (file names and line numbers) from the messages. This is useful to make tests reproducible.