</config>
A last solution is to pass your configuration directly in your program
-with :cpp:func:`simgrid::s4u::Engine::set_config` or :cpp:func:`MSG_config`.
+with :cpp:func:`simgrid::s4u::Engine::set_config`.
.. code-block:: cpp
- **host/model:** :ref:`options_model_select`
-- **maxmin/precision:** :ref:`cfg=maxmin/precision`
- **maxmin/concurrency-limit:** :ref:`cfg=maxmin/concurrency-limit`
-- **msg/debug-multiple-use:** :ref:`cfg=msg/debug-multiple-use`
-
- **model-check:** :ref:`options_modelchecking`
- **model-check/checkpoint:** :ref:`cfg=model-check/checkpoint`
- **model-check/communications-determinism:** :ref:`cfg=model-check/communications-determinism`
- **model-check/reduction:** :ref:`cfg=model-check/reduction`
- **model-check/replay:** :ref:`cfg=model-check/replay`
- **model-check/send-determinism:** :ref:`cfg=model-check/send-determinism`
+- **model-check/setenv:** :ref:`cfg=model-check/setenv`
- **model-check/termination:** :ref:`cfg=model-check/termination`
- **model-check/timeout:** :ref:`cfg=model-check/timeout`
- **model-check/visited:** :ref:`cfg=model-check/visited`
- **storage/max_file_descriptors:** :ref:`cfg=storage/max_file_descriptors`
-- **surf/precision:** :ref:`cfg=surf/precision`
+- **precision/timing:** :ref:`cfg=precision/timing`
+- **precision/work-amount:** :ref:`cfg=precision/work-amount`
- **For collective operations of SMPI,** please refer to Section :ref:`cfg=smpi/coll-selector`
- **smpi/auto-shared-malloc-thresh:** :ref:`cfg=smpi/auto-shared-malloc-thresh`
- **smpi/async-small-thresh:** :ref:`cfg=smpi/async-small-thresh`
+- **smpi/barrier-finalization:** :ref:`cfg=smpi/barrier-finalization`
+- **smpi/barrier-collectives:** :ref:`cfg=smpi/barrier-collectives`
- **smpi/buffering:** :ref:`cfg=smpi/buffering`
-- **smpi/bw-factor:** :ref:`cfg=smpi/bw-factor`
- **smpi/coll-selector:** :ref:`cfg=smpi/coll-selector`
- **smpi/comp-adjustment-file:** :ref:`cfg=smpi/comp-adjustment-file`
- **smpi/cpu-threshold:** :ref:`cfg=smpi/cpu-threshold`
- **smpi/display-allocs:** :ref:`cfg=smpi/display-allocs`
- **smpi/display-timing:** :ref:`cfg=smpi/display-timing`
- **smpi/errors-are-fatal:** :ref:`cfg=smpi/errors-are-fatal`
-- **smpi/finalization-barrier:** :ref:`cfg=smpi/finalization-barrier`
- **smpi/grow-injected-times:** :ref:`cfg=smpi/grow-injected-times`
- **smpi/host-speed:** :ref:`cfg=smpi/host-speed`
- **smpi/IB-penalty-factors:** :ref:`cfg=smpi/IB-penalty-factors`
- **smpi/iprobe-cpu-usage:** :ref:`cfg=smpi/iprobe-cpu-usage`
- **smpi/init:** :ref:`cfg=smpi/init`
- **smpi/keep-temps:** :ref:`cfg=smpi/keep-temps`
-- **smpi/lat-factor:** :ref:`cfg=smpi/lat-factor`
- **smpi/ois:** :ref:`cfg=smpi/ois`
- **smpi/or:** :ref:`cfg=smpi/or`
- **smpi/os:** :ref:`cfg=smpi/os`
<https://hal.inria.fr/inria-00071989/document>`_.
- **ns-3** (only available if you compiled SimGrid accordingly):
Use the packet-level network
- simulators as network models (see :ref:`model_ns3`).
+ simulators as network models (see :ref:`models_ns3`).
This model can be :ref:`further configured <options_pls>`.
-- ``cpu/model``: specify the used CPU model. We have only one model
- for now:
+- ``cpu/model``: specify the used CPU model. We have only one model for now:
- **Cas01:** Simplistic CPU model (time=size/speed)
-- ``host/model``: The host concept is the aggregation of a CPU with a
- network card. Three models exists, but actually, only 2 of them are
- interesting. The "compound" one is simply due to the way our
- internal code is organized, and can easily be ignored. So at the
- end, you have two host models: The default one allows aggregation of
- an existing CPU model with an existing network model, but does not
- allow parallel tasks because these beasts need some collaboration
- between the network and CPU model.
-
- - **default:** Default host model. Currently, CPU:Cas01 and
- network:LV08 (with cross traffic enabled)
- - **compound:** Host model that is automatically chosen if
- you change the network and CPU models
- - **ptask_L07:** Host model somehow similar to Cas01+CM02 but
- allowing "parallel tasks", that are intended to model the moldable
- tasks of the grid scheduling literature.
+- ``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.
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:
and slow pattern that follows the actual dependencies.
.. _cfg=bmf/precision:
-.. _cfg=maxmin/precision:
-.. _cfg=surf/precision:
+.. _cfg=precision/timing:
+.. _cfg=precision/work-amount:
Numerical Precision
...................
-**Option** ``maxmin/precision`` **Default:** 1e-5 (in flops or bytes) |br|
-**Option** ``surf/precision`` **Default:** 1e-9 (in seconds) |br|
+**Option** ``precision/timing`` **Default:** 1e-9 (in seconds) |br|
+**Option** ``precision/work-amount`` **Default:** 1e-5 (in flops or bytes) |br|
**Option** ``bmf/precision`` **Default:** 1e-12 (no unit)
The analytical models handle a lot of floating point values. It is
**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. 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.
+The analytical models need to know the maximal TCP window size to take the TCP congestion mechanism into account (see
+:ref:`this page <understanding_cm02>` 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.
.. code-block:: console
$ 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.
+
.. _cfg=network/bandwidth-factor:
.. _cfg=network/latency-factor:
.. _cfg=network/weight-S:
-Correcting Important Network Parameters
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-SimGrid can take network irregularities such as a slow startup or
-changing behavior depending on the message size into account. You
-should not change these values unless you really know what you're
-doing. The corresponding values were computed through data fitting
-one the timings of packet-level simulators, as described in `Accuracy
-Study and Improvement of Network Simulation in the SimGrid Framework
-<http://mescal.imag.fr/membres/arnaud.legrand/articles/simutools09.pdf>`_.
-
-- **network/latency-factor**: apply a multiplier to latency.
- Models the TCP slow-start mechanism.
-- **network/bandwidth-factor**: actual bandwidth perceived by the
- user.
-- **network/weight-S**: bottleneck sharing constant parameter. Used
- to calculate RTT.
-
-These parameters are the same for all communications in your simulation,
-independently of message size or source/destination hosts. A more flexible
-mechanism based on callbacks was introduced in SimGrid. It provides the user
-a callback that will be called for each communication, allowing the user
-to set different latency and bandwidth factors, based on the message size, links used
-or zones traversed. To more details of how to use it, please look at the
-`examples/cpp/network-factors/s4u-network-factors.cpp <https://framagit.org/simgrid/simgrid/tree/master/examples/cpp/network-factors/s4u-network-factors.cpp>`_.
-
-
-If you are using the SMPI model, these correction coefficients are
-themselves corrected by constant values depending on the size of the
-exchange. By default SMPI uses factors computed on the Stampede
-Supercomputer at TACC, with optimal deployment of processes on
-nodes. Again, only hardcore experts should bother about this fact.
-For more details, see SMPI sections about :ref:`cfg=smpi/bw-factor` and :ref:`cfg=smpi/lat-factor`.
+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 :ref:`in the relevant chapter <models_calibration>`.
+
+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
+:ref:`network path <platform_routing>` 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
+<https://framagit.org/simgrid/simgrid/tree/master/examples/cpp/network-factors/s4u-network-factors.cpp>`_. 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
+:ref:`calibration <models_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 :ref:`models_TCP` and also this scientific paper: `Accuracy Study and Improvement of Network
+Simulation in the SimGrid Framework <http://mescal.imag.fr/membres/arnaud.legrand/articles/simutools09.pdf>`_
+Default values for ``CM02`` is 0. ``LV08`` sets it to 20537 while both ``SMPI`` and ``IB`` set it to 8775.
+
+.. _cfg=network/loopback:
+
+Configuring loopback link
+^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Several network models provide an implicit loopback link to account for local
+communication on a host. By default it has a 10GBps bandwidth and a null latency.
+This can be changed with ``network/loopback-lat`` and ``network/loopback-bw``
+items. Note that this loopback is conveniently modeled with a :ref:`single FATPIPE link <pf_loopback>`
+for the whole platform. If modeling contention inside nodes is important then you should
+rather add such loopback links (one for each host) yourself.
.. _cfg=smpi/IB-penalty-factors:
^^^^^^^^^^^^^^^^
InfiniBand network behavior can be modeled through 3 parameters
-``smpi/IB-penalty-factors:"βe;βs;γs"``, as explained in `this PhD
-thesis
+``smpi/IB-penalty-factors:"βe;βs;γs"``, as explained in `the PhD
+thesis of Jérôme Vienne
<http://mescal.imag.fr/membres/jean-marc.vincent/index.html/PhD/Vienne.pdf>`_ (in French)
or more concisely in `this paper <https://hal.inria.fr/hal-00953618/document>`_,
even if that paper does only describe models for myrinet and ethernet.
Note that with the default host model this option is activated by default.
-.. _cfg=network/loopback:
-
-Configuring loopback link
-^^^^^^^^^^^^^^^^^^^^^^^^^
-
-Several network model provide an implicit loopback link to account for local
-communication on a host. By default it has a 10GBps bandwidth and a null latency.
-This can be changed with ``network/loopback-lat`` and ``network/loopback-bw``
-items.
-
.. _cfg=smpi/async-small-thresh:
Simulating Asynchronous Send
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
-:cpp:func:`MSG_mailbox_set_async`. After this, all messages sent to
+:cpp:func:`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
communication determinism mode of the model checker, which checks
determinism properties of the communications of an application.
+.. _cfg=model-check/setenv:
+
+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``.
+
.. _options_mc_perf:
Verification Performance Considerations
auto-detection fails for you). They are approximately sorted here from
the slowest to the most efficient:
- - **thread:** very slow factory using full featured threads (either
- pthreads or windows native threads). They are slow but very
- standard. Some debuggers or profilers only work with this factory.
- - **java:** Java applications are virtualized onto java threads (that
- are regular pthreads registered to the JVM)
+ - **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 <http://www.boost.org/doc/libs/1_59_0/libs/context/doc/html/index.html>`_
Disabling Stack Guard Pages
...........................
-**Option** ``contexts/guard-size`` **Default** 1 page in most case (0 pages on Windows or with MC)
+**Option** ``contexts/guard-size`` **Default** 1 page in most case (0 pages with MC)
Unless you use the threads context factory (see
:ref:`cfg=contexts/factory`), a stack guard page is usually used
.............................
Parallel execution of the user code is only considered stable in
-SimGrid v3.7 and higher, and mostly for MSG simulations. SMPI
+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 <http://hal.inria.fr/inria-00602216/>`_.
you never used the tracing API.
-- Any SimGrid-based simulator (MSG, SMPI, ...) and raw traces:
+- Any SimGrid-based simulator (S4U, SMPI, ...) and raw traces:
.. code-block:: none
tells it to trace host and link utilization (without any
categorization).
-- MSG-based simulator and categorized traces (you need to
+- S4U-based simulator and categorized traces (you need to
declare categories and classify your tasks according to them)
.. code-block:: none
simulations. For additional details about this and all tracing
options, check See the :ref:`tracing_tracing_options`.
-Configuring MSG
----------------
-
-.. _cfg=msg/debug-multiple-use:
-
-Debugging MSG Code
-..................
-
-**Option** ``msg/debug-multiple-use`` **Default:** off
-
-Sometimes your application may try to send a task that is still being
-executed somewhere else, making it impossible to send this task. However,
-for debugging purposes, one may want to know what the other host is/was
-doing. This option shows a backtrace of the other process.
-
Configuring SMPI
----------------
The SMPI interface provides several specific configuration items.
-These are not easy to see, since the code is usually launched through the
-``smiprun`` script directly.
+These are not easy to see with ``--help-cfg``, since SMPI binaries are usually launched through the ``smiprun`` script.
.. _cfg=smpi/host-speed:
.. _cfg=smpi/cpu-threshold:
smpicc or smpiff, respectively. This flag activates some internal
macro definitions that help with obtaining the call location.
-.. _cfg=smpi/bw-factor:
-
-Bandwidth Factors
-.................
-
-**Option** ``smpi/bw-factor``
-|br| **Default:** 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
-
-The possible throughput of network links is often dependent on the
-message sizes, as protocols may adapt to different message sizes. With
-this option, a series of message sizes and factors are given, helping
-the simulation to be more realistic. For instance, the current default
-value means that messages with size 65472 bytes and more will get a total of
-MAX_BANDWIDTH*0.940694, messages of size 15424 to 65471 will get
-MAX_BANDWIDTH*0.697866, and so on (where MAX_BANDWIDTH denotes the
-bandwidth of the link).
+Bandwidth and latency factors
+.............................
-An experimental script to compute these factors is available online. See
-https://framagit.org/simgrid/platform-calibration/
-https://simgrid.org/contrib/smpi-saturation-doc.html
+Adapting the bandwidth and latency acurately to the network conditions is of a paramount importance to get realistic results.
+This is done through the :ref:`network/bandwidth-factor <cfg=network/bandwidth-factor>` and :ref:`network/latency-factor
+<cfg=network/latency-factor>` items. You probably also want to read the following section: :ref:`models_calibration`.
.. _cfg=smpi/display-timing:
under the dlopen privatization schema, which tends to fool the
debuggers.
-.. _cfg=smpi/lat-factor:
-
-Latency factors
-...............
-
-**Option** ``smpi/lat-factor`` |br|
-**default:** 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
-
-The motivation and syntax for this option is identical to the motivation/syntax
-of :ref:`cfg=smpi/bw-factor`.
-
-There is an important difference, though: While smpi/bw-factor `reduces` the
-actual bandwidth (i.e., values between 0 and 1 are valid), latency factors
-increase the latency, i.e., values larger than or equal to 1 are valid here.
-
.. _cfg=smpi/papi-events:
Trace hardware counters with PAPI
.. TODO:: All available collective algorithms will be made available
via the ``smpirun --help-coll`` command.
-.. _cfg=smpi/finalization-barrier:
+.. _cfg=smpi/barrier-collectives:
+
+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.
+
+.. code-block:: C
+
+ 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.
+
+.. _cfg=smpi/barrier-finalization:
Add a barrier in MPI_Finalize
.............................
