Skip to content

Latest commit

 

History

History
66 lines (52 loc) · 3.64 KB

README_Services.md

File metadata and controls

66 lines (52 loc) · 3.64 KB
Raw

FAODEL-Services: Simple Services for FAODEL

FAODEL provides a few optional services that can be launched during the bootstrap process:

BackBurner

Many communication services need a simple way to queue up tasks that need to happen at a later point in time (e.g., notifying clients a new object is available, implementing a timeout check, queueing communication operations to avoid deadlock). BackBurner is a generic progress thread developers can use to queue up work. BackBurner provides a fifo queue for handing off operations and can be configured to use multiple threads (each with its own work queue). The BackBurner worker thread pulls a bundle of queue operations out at a time and then processes each one in order.

BackBurner has a few mechanisms for controlling how workers are notified of new work:

  • Pipe (default): A pipe is used to pass notifications of new work to the workers. While this increases latencies, it allows the kernel to make a decision about when to schedule the worker thread.
  • Polling: Workers continuously check their queue sizes. This method improve latencies, but it wastes a tremendous amount of CPU time.
  • Sleep Polling: Workers do a sleep operation before they poll for new work. Users can adjust the polling rate by changing the sleep time. Given that a worker still processes all the entries that are available when it wakes up, this approach can be useful for batching up work.

BackBurner parses the Configuration object for the following settings:

Property Type Default Description
backburner.debug boolean false Display debug messages
backburner.threads integer 1 Number of worker threads to use
backburner.notification_method pipe, polling, sleep_polling pipe Controls how workers are notified of new work
backburner.sleep_polling_time time (us) 100us Set polling delay in sleep_polling mode

MPISyncStart

FAODEL applications often need a simple way to pass initial configuration information between nodes. The MPISyncStart service examines a configuration and updates certain properties (ending with "_mpi") with runtime information.

For properties that allow a user to specify a list of nodes (eg dirman.resources_mpi), the parser supports a few ways to build node lists:

  • Rank IDs: The actual rank numbers, such as 0 or 0,1,2 or 0-3
  • "middle": the middle rank. eg in an 8 rank job, 0-middle would be the first half
  • "end": the last rank in the job. eg in an 8 rank job, end would be 7.
  • "all": all of the ranks, eg in an 8 rank job all would be 0-7.

MPISyncStart parses the Configuration object for the following settings:

Property Type Default Description
mpisyncstart.debug boolean false Display debug messages
mpisyncstart.enable boolean false Enable the service
dirman.root_node_mpi Rank -1 Specify rank for dirman root node
dirman.resources_mpi[] url node(s) "" Specify resource using mpi ranks

Note: Bootstrap has an mpisyncstop.enable option that performs a barrier on shutdown.