Architecture

Mars provides a library for distributed execution of tensors. The distributed applications are built with actor model provided by mars.actors and consists of three parts: the scheduler, the worker and the web service.

Users submit their tasks in graphs built with tensors. The web service accepts tensor graphs and sends them into a scheduler, where graphs are compiled into operand graphs, analyzed and partitioned before submitted to workers. The scheduler then creates and scatters operand actors who control task execution on workers on other schedulers given consistent hashing. Then operands are activated and executed in topological order. When all operands related to terminating tensors are executed, the graph will be marked as finished and the client can pull the result from workers, proxied by the scheduler. The whole procedure can be seen in the graph below.

../_images/mars-exec-flow.svg

Job Submission

Jobs are submitted into Mars via RESTful APIs. Users type tensor operations and run a tensor by calling session.run(tensor), which builds a tensor graph given the operations created by the user. This graph is sent to the web api and a GraphActor is created given consistent hashing in the cluster to handle the tensor graph. After that the web client begins querying the state of the graph until termination.

In the GraphActor, we first convert the tensor graph into an operand graph via tiling methods. This enables the graph to run in parallel. After that, several analyzes are performed on the graph to obtain operand priorities and assign workers for the operand, which can be seen in detail in graph preparation and scheduling policy section. Then OperandActor is created for every operand to control detailed execution. When an operand is in READY state, as described in operand states section, a worker will be selected and the operand is submitted into the worker for execution.

Execution Control

When an operand is submitted to a worker, the OperandActor on the scheduler listen to its callback. When the execution is successful, successors of that operand will be scheduled. When the execution failed, the OperandActor will retry several times before announcing the execution as fatal.

Job Cancellation

Users can cancel a running job via RESTful API. The request is written into state storage first and then called in GraphActor. If the graph is under preparation, it will stop immediately when the stop request is detected in state storage. Otherwise every operand is scanned and the states will be set as CANCELLING. When the operand is currently running, a stop request will be sent into workers which results in ExecutionInterrupted exception in workers. When this exception is received in OperandActor, the state of the operand will be marked as CANCELLED.