WRENCH 102

In WRENCH’s terminology, and execution controller is software that makes all decisions and takes all actions for executing some application workflow using cyberinfrastructure services. It is thus a crucial component in every WRENCH simulator. WRENCH does not provide any execution controller implementation, but provides the means for developing custom ones. This page is meant to provide high-level and detailed information about implementing an execution controller in WRENCH. Full API details are provided in the Developer API Reference.

Basic blueprint for an execution controller implementation

An execution controller implementation needs to use many WRENCH classes, which are accessed by including a single header file:

#include <wrench-dev.h>

An execution controller implementation must derive the wrench::ExecutionController class, which means that it must override several the virtual main() member function. A typical such implementation of this function goes through a simple loop as follows:

// A) create/retrieve application workload to execute
// B) obtain information about running services
while (application workload execution has not completed/failed) {
  // C) interact with services
  // D) wait for an event and react to it
}

In the next three sections, we give details on how to implement the above. To provide context, we make frequent references to the execution controllers implemented as part of the example simulators in the examples/ directory. Afterwards are a few sections that highlight features and functionality relevant to execution controller development.

A) Finding out information about running services

Services that the execution controller can use are typically passed to its constructor. Most service classes provide member functions to get information about the capabilities and properties of the services. For instance, a wrench::ComputeService has a wrench::ComputeService::getNumHosts() member function that returns how many compute hosts the service has access to in total. A wrench::StorageService has a wrench::StorageService::getFreeSpace() member function to find out how many bytes of free space are available on it. And so on…

To take a concrete example, consider the execution controller implementation in examples/workflow_api/basic-examples/batch-bag-of-tasks/TwoTasksAtATimeBatchWMS.cpp. This WMS finds out the compute speed of the cores of the compute nodes available to a wrench::BatchComputeService as:

double core_flop_rate = (*(batch_service->getCoreFlopRate().begin())).second;

Member function wrench::ComputeService::getCoreFlopRate() returns a map of core compute speeds indexed by hostname (the map thus has one element per compute node available to the service). Since the compute nodes of a batch compute service are homogeneous, the above code simply grabs the core speed value of the first element in the map.

It is important to note that these member functions actually involve communication with the service, and thus incur overhead that is part of the simulation (as if, in the real-world, you would contact a running service with a request for information over the network). This is why the line of code above, in that example execution controller, is executed once and the core compute speed is stored in the core_flop_rate variable to be re-used by the execution controller repeatedly throughout its execution.

B) Interacting with services

An execution controller can have many and complex interactions with services, especially with compute and storage services. In this section, we describe how WRENCH makes these interactions relatively easy, providing examples for each kind of interaction for each kind of service.

Job Manager and Data Movement Manager

As expected, each service type provides its own API. For instance, a network proximity service provides member functions to query the service’s host distance databases. The Developer API Reference provides all necessary documentation, which also explains which member functions are synchronous and which are asynchronous (in which case some event will occur in the future). However, the WRENCH developer will find that many member functions that one would expect are nowhere to be found. For instance, the compute services do not have (public) member functions for submitting jobs for execution!

The rationale for the above is that many member functions need to be asynchronous so that the execution controller can use services concurrently. For instance, an execution controller could submit a job to two distinct compute services asynchronously, and then wait for the service which completes its job first and cancel the job on the other service. Exposing this asynchronicity to the execution controller would require that the WRENCH developer use data structures to perform the necessary bookkeeping of ongoing service interactions, and process incoming control messages from the services on the (simulated) network or alternately register many callbacks. Instead, WRENCH provides managers. One can think of managers as separate threads that handle all asynchronous interactions with services, and which have been implemented for your convenience to make interacting with services easy.

There are two managers: a job manager (class wrench::JobManager) and a data movement manager (class wrench::DataMovementManager). The base wrench::ExecutionController class provides two member functions for instantiating and starting these managers: wrench::ExecutionController::createJobManager() and wrench::ExecutionController::createDataMovementManager().

Creating one or two of these managers typically is the first thing an execution controller does. For instance, the execution controller in examples/workflow_api/basic-examples/bare-metal-data-movement/DataMovementWMS.cpp starts by doing:

auto job_manager = this->createJobManager();
auto data_movement_manager = this->createDataMovementManager();

Each manager has its own documented API, and is discussed further in sections below.

Interacting with storage services

Typical interactions between an execution controller and a storage service include locating, reading, writing, and copying files. Different storage service implementations may or not implement some of of these operations. Click on the following links to see concrete examples of interactions with the currently available storage service type:

Interacting with compute services

The Job abstraction

The main activity of an execution controller is to execute workflow tasks on compute services. Rather than submitting tasks directly to compute services, an execution controller must create “jobs”, which can comprise multiple tasks and involve data copy/deletion operations. The job abstraction is powerful and greatly simplifies the task of an execution controller while affording flexibility.

There are three kinds of jobs in WRENCH: wrench::CompoundJob, wrench::StandardJob, and wrench::PilotJob.

A Compound Job is simply set of actions to be performed, with possible control dependencies between actions. It is the most generic, flexible, and expressive kind of job. See the API documentation for the wrench::CompoundJob class and the examples in the examples/action_api directory. The other types of jobs below are actually implemented internally as compound jobs. The Compound Job abstraction is the most recent addition to the WRENCH API, and vastly expands the list of possible things that an execution controller can do. But because it is more recent, the reader will find that there are more examples in these documents and in the examples directory that use standard jobs (described below). But all these examples could be easily rewritten using the more generic compound job abstraction.

A Standard Job is a specific kind of job designed for workflow applications. In its most complete form, a standard job specifies: - A set (in fact a vector) of std::shared_ptr<wrench::WorkflowTask> to execute, so that each task without all its predecessors in the set is ready;

  • A std::map of <std::shared_ptr<wrench::DataFile>>, std::shared_ptr<wrench::StorageService>> pairs that specifies from which storage services particular input files should be read and to which storage services output files should be written;

  • A set of file copy operations to be performed before executing the tasks;

  • A set of file copy operations to be performed after executing the tasks; and

  • A set of file deletion operations to be performed after executing the tasks and file copy operations.

Any of the above can actually be empty, and in the extreme a standard job can do nothing.

A Pilot Job (sometimes called a “placeholder job” in the literature) is a concept that is mostly relevant for batch scheduling. In a nutshell, it is a job that allows late binding of tasks to resources. It is submitted to a compute service (provided that service supports pilot jobs), and when it starts it just looks to the execution controller like a short-lived wrench::BareMetalComputeService to which compound and/or standard jobs can be submitted.

All jobs are created via the job manager, which provides wrench::JobManager::createCompoundJob(), wrench::JobManager::createStandardJob(), and wrench::JobManager::createPilotJob() member functions (the job manager is thus a job factory).

In addition to member functions for job creation, the job manager also provides the following:

  • wrench::JobManager::submitJob(): asynchronous submission of a job to a compute service.

  • wrench::JobManager::terminateJob(): synchronous termination of a previously submitted job.

The next section gives examples of interactions with each kind of compute service.

Click on the following links to see detailed descriptions and examples of how jobs are submitted to each compute service type:

Interacting with file registry services

Interaction with a file registry service is straightforward and done by directly calling member functions of the wrench::FileRegistryService class. Note that often file registry service entries are managed automatically, e.g., via calls to wrench::DataMovementManager and wrench::StorageService member functions. So often an execution controller does not need to interact with the file registry service.

Adding/removing an entry to a file registry service is done as follows:

std::shared_ptr<wrench::FileRegistryService> file_registry;
std::shared_ptr<wrench::DataFile> some_file;
std::shared_ptr<wrench::StorageService> some_storage_service;

[...]

file_registry->addEntry(wrench::FileLocation::LOCATION(some_storage_service, some_file));
file_registry->removeEntry(wrench::FileLocatio::LOCATION(some_storage_service, some_file));

The wrench::FileLocation class is a convenient abstraction for a file that is available at some storage service (with optionally a directory path at that service).

Retrieving all entries for a given file is done as follows:

std::shared_ptr<wrench::FileRegistryService> file_registry;
std::shared_ptr<wrench::DataFile> some_file;

[...]

std::set<std::shared_ptr<wrench::FileLocation>> entries;
entries = file_registry->lookupEntry(some_file);

If a network proximity service is running, it is possible to retrieve entries for a file sorted by non-decreasing proximity from some reference host. Returned entries are stored in a (sorted) std::map where the keys are network distances to the reference host. For instance:

std::shared_ptr<wrench::FileRegistryService> file_registry;
std::shared_ptr<wrench::DataFile> some_file;
std::shared_ptr<wrench::NetworkProximityService> np_service;

[...]

auto entries = fr_service->lookupEntry(some_file, "ReferenceHost", np_service);

See the documentation of wrench::FileRegistryService for more API member functions.

Interacting with network proximity services

Querying a network proximity service is straightforward. For instance, to obtain a measure of the network distance between hosts “Host1” and “Host2”, one simply does:

std::shared_ptr<wrench::NetworkProximityService> np_service;

double distance = np_service->query(std::make_pair("Host1","Host2"));

This distance corresponds to half the round-trip-time, in seconds, between the two hosts. If the service is configured to use the Vivaldi coordinate-based system, as in our example above, this distance is actually derived from network coordinates, as computed by the Vivaldi algorithm. In this case, one can actually ask for these coordinates for any given host:

std::pair<double,double> coords = np_service->getCoordinates("Host1");

See the documentation of wrench::NetworkProximityService for more API member functions.

C) Workflow execution events

Because the execution controller performs asynchronous operations, it needs to wait for and re-act to events. This is done by calling the wrench::ExecutionController::waitForAndProcessNextEvent() member function implemented by the base wrench::ExecutionController class. A call to this member function blocks until some event occurs and then calls a callback member function. The possible event classes all derive from the wrench::ExecutionEvent class, and an execution controller can override the callback member function for each possible event (the default member function does nothing but print some log message). These overridable callback member functions are:

  • wrench::ExecutionController::processEventCompoundJobCompletion(): react to a compound job completion

  • wrench::ExecutionController::processEventCompoundJobFailure(): react to a compound job failure

  • wrench::ExecutionController::processEventStandardJobCompletion(): react to a standard job completion

  • wrench::ExecutionController::processEventStandardJobFailure(): react to a standard job failure

  • wrench::ExecutionController::processEventPilotJobStart(): react to a pilot job beginning execution

  • wrench::ExecutionController::processEventPilotJobExpiration(): react to a pilot job expiration

  • wrench::ExecutionController::processEventFileCopyCompletion(): react to a file copy completion

  • wrench::ExecutionController::processEventFileCopyFailure(): react to a file copy failure

Each member function above takes in an event object as parameter. In the case of failure, the event includes a wrench::FailureCause object, which can be accessed to analyze (or just display) the root cause of the failure.

Consider the execution controller in examples/workflow_api/basic-examples/bare-metal-bag-of-tasks/TwoTasksAtATimeWMS.cpp. At each each iteration of its main loop it does:

// Submit some standard job to some compute  service
job_manager->submitJob(...);

// Wait for and process next event
this->waitForAndProcessNextEvent();

In this simple example, only one of two events could occur at this point: a standard job completion or a standard job failure. As a result, this execution controller overrides the two corresponding member functions as follows:

void TwoTasksAtATimeWMS::processEventStandardJobCompletion(
               std::shared_ptr<StandardJobCompletedEvent> event) {
  // Retrieve the job that this event is for
  auto job = event->job;
  // Print some message for each task in the job
  for (auto const &task : job->getTasks()) {
    std::cerr  << "Notified that a standard job has completed task " << task->getID() << std::endl;
  }
}

void TwoTasksAtATimeWMS::processEventStandardJobFailure(
               std::shared_ptr<StandardJobFailedEvent> event) {
  // Retrieve the job that this event is for
  auto job = event->job;
  std::cerr  << "Notified that a standard job has failed (failure cause: ";
  std::cerr << event->failure_cause->toString() << ")" <<  std::endl;
  // Print some message for each task in the job if it has failed
  std::cerr << "As a result, the following tasks have failed:";
  for (auto const &task : job->getTasks()) {
    if (task->getState != WorkflowTask::COMPLETE) {
      std::cerr  << "  - " << task->getID() << std::endl;
    }
  }
}

You may note some difference between the above code and that in examples/workflow_api/basic-examples/bare-metal-bag-of-tasks/TwoTasksAtATimeWMS.cpp. This is for clarity purposes, and especially because we have not yet explained how WRENCH does message logging. See an upcoming section about logging.

While the above callbacks are convenient, sometimes it is desirable to do things more manually. That is, wait for an event and then process it in the code of the main loop of the execution controller rather than in a callback member function. This is done by calling the wrench::waitForNextEvent() member function. For instance, the execution controller in examples/workflow_api/basic-examples/bare-metal-data-movement/DataMovementWMS.cpp does it as:

// Initiate an asynchronous file copy
data_movement_manager->initiateAsynchronousFileCopy(...);

// Wait for an event
auto event = this->waitForNextEvent();

//Process the event
if (auto file_copy_completion_event = std::dynamic_pointer_cast<wrench::FileCopyCompletedEvent>(event)) {
  std::cerr << "Notified of a file copy completion for file ";
  std::cerr << file_copy_completion_event->file->getID()<< "as expected" << std::endl;
} else {
   throw std::runtime_error("Unexpected event (" + event->toString() + ")");}
}

Exceptions

Most member functions in the WRENCH Developer API throw exceptions. In fact, most of the code fragments above should be in try-catch clauses, catching these exceptions.

Some exceptions correspond to failures during the simulated workflow executions (i.e., errors that would occur in a real-world execution and are thus part of the simulation). Each such exception contains a wrench::FailureCause object, which can be accessed to understand the root cause of the execution failure. Other exceptions (e.g., std::invalid_arguments, std::runtime_error) are thrown as well, which are used for detecting misuses of the WRENCH API or internal WRENCH errors.

Finding information and interacting with hardware resources

The wrench::Simulation class provides many member functions to discover information about the (simulated) hardware platform and interact with it. It also provides other useful information about the simulation itself, such as the current simulation date. Some of these member functions are static, but others are not. The wrench::ExecutionController class includes a simulation object. Thus, the execution controller can call member functions on the this->simulation object. For instance, this fragment of code shows how an execution controller can figure out the current simulated date and then check that a host exists (given a hostname) and, if so, set its pstate (power state) to the highest possible setting.

auto now = wrench::Simulation::getCurrentSimulatedDate();
if (wrench::Simulation::doesHostExist("SomeHost"))  {
  this->simulation->setPstate("SomeHost", wrench::Simulation::getNumberofPstates("SomeHost")-1);
}

See the documentation of the wrench::Simulation class for all details. Specifically regarding host pstates, see the example execution controller in examples/workflow_api/basic-examples/cloud-bag-of-tasks-energy/TwoTasksAtATimeCloudWMS.cpp, which interacts with host pstates (and the examples/workflow_api/basic-examples/cloud-bag-of-tasks-energy/four_hosts_energy.xml platform description file which defines pstates).

Logging

It is typically desirable for the execution controller to print log output to the terminal. This is easily accomplished using the wrench::WRENCH_INFO(), wrench::WRENCH_DEBUG(), and wrench::WRENCH_WARN() macros, which are used just like C’s printf(). Each of these macros corresponds to a different logging level in SimGrid. See the SimGrid logging documentation for all details.

Furthermore, one can change the color of the log messages with the wrench::TerminalOutput::setThisProcessLoggingColor() member function, which takes as parameter a color specification:

  • wrench::TerminalOutput::COLOR_BLACK

  • wrench::TerminalOutput::COLOR_RED

  • wrench::TerminalOutput::COLOR_GREEN

  • wrench::TerminalOutput::COLOR_YELLOW

  • wrench::TerminalOutput::COLOR_BLUE

  • wrench::TerminalOutput::COLOR_MAGENTA

  • wrench::TerminalOutput::COLOR_CYAN

  • wrench::TerminalOutput::COLOR_WHITE

When inspecting the code of the execution controllers in the example simulators you will find many examples of calls to wrench::WRENCH_INFO(). The logging is per .cpp file, each of which corresponds to a declared logging category. For instance, in examples/workflow_api/basic-examples/batch-bag-of-tasks/TwoTasksAtATimeBatchWMS.cpp, you will find the typical pattern:

// Define a log category name for this file
WRENCH_LOG_CATEGORY(custom_wms, "Log category for TwoTasksAtATimeBatchWMS");

[...]

int TwoTasksAtATimeBatchWMS::main() {

  // Set the logging color to green
  TerminalOutput::setThisProcessLoggingColor(TerminalOutput::COLOR_GREEN);

  [...]

  // Print an info-level message, using printf-like format
  WRENCH_INFO("Submitting the job, asking for %s %s-core nodes for %s minutes",
              service_specific_arguments["-N"].c_str(),
              service_specific_arguments["-c"].c_str(),
              service_specific_arguments["-t"].c_str());

  [...]

  // Print a last info-level message
  WRENCH_INFO("Workflow execution complete");
  return 0;
}

The name of the logging category, in this case custom_wms, can then be passed to the --log command-line argument. For instance, invoking the simulator with additional argument --log=custom_wms.threshold=info will make it so that only those WRENCH_INFO statements in TwoTasksAtATimeBatchWMS.cpp will be printed (in green!).