Loss

The story

A loss is final. Once a device is lost, rig counts as done only the values the device had reached when the loss was recorded. The work of later values may have run, in part or not at all, and no program can tell which, so rig refuses every use whose validity depends on that work. Every use of the lost device and of its memory raises Rig.Lost. Other memory raises it only for the uses that depend on a point the lost device never reached, and the Loss section lists what each use depends on. The worked trace shows both a buffer that stays usable and one that raises for the rest of the process.

Other devices go on, unless their queues wait for a value the lost device never reached. A lost device's stop runs only when the program closes the device or opens its name again, never inside the call that met the loss, so a program that catches Rig.Lost closes the device. Until its stop the device keeps its memory and the values its submissions keep, and a GPU driven over PCI may go on running the work it took. Once the device has stopped, as the Stops section of rig.mli defines it, which may be never, its memory returns, and memory it only read is ordinary again. Memory whose last write it never reached raises Rig.Lost for the rest of the process.

What a program sees

Rig.Lost (d, why) carries the device and its reason, and Rig.lost d returns the same reason without raising. The reasons are listed at Rig.Lost. The process's failure, Rig.failure, is its first loss that Rig.close did not cause, or the reason given to Rig.fail. A job that runs on several machines fails as a whole (Rig_remote).

Closing a device loses it with the reason "closed", and opening its name again makes a new device:

# open Rig;;

# let () =
    let a = Result.get_ok (memory_device "A") in
    let b = Result.get_ok (memory_device "B") in
    let x = Buffer.create a 4 in
    close a;
    (match Buffer.copy ~src:x ~dst:(Buffer.create b 4) with
     | () -> print_endline "copied"
     | exception (Lost _ as e) -> print_endline (Printexc.to_string e));
    Printf.printf "lost: %s, failure: %b\n"
      (Option.value (lost a) ~default:"no")
      (Option.is_some (failure ()));
    let a' = Result.get_ok (memory_device "A") in
    Printf.printf "A again: same device %b, submitted %d\n"
      (equal a a') (submitted a');;

A lost: closed
lost: closed, failure: false
A again: same device false, submitted 0

this page is tested with mdxedit this page ↗