Compound applications

Compound applications

A compound application consists of a main program and one or more worker programs. The main program executes the worker programs in sequence, and maintains a 'main state file' that records which worker programs have completed. The main program assigns to each worker program a subrange of the overall 'fraction done' range of 0..1. For example, if there are two worker programs with equal runtime, the first would have range 0 to 0.5 and the second would have range 0.5 to 1. The BOINC API provides a number of functions, and in developing a compound application you must decide whether these functions are to be performed by the main or worker program. The functions are represented by flags in the BOINC_OPTIONS structure. Each flag must be set in either the main or worker programs, but not both.

struct BOINC_OPTIONS {
    int main_program;
    int check_heartbeat;
    int handle_trickle_ups;
    int handle_trickle_downs;
    int handle_process_control;
    int handle send_status_msgs;
    int direct_process_action;
};

int boinc_init_options(BOINC_OPTIONS*);

main_program

Set this in the main program.

check_heartbeat

If set, the program monitors 'heartbeat' messages from the core client. If the heartbeat stops, the result depends on the direct_process_action flag (see below).

handle_trickle_ups

If set, the program can send trickle-up messages.

handle_trickle_downs

If set, the program can receive trickle-down messages.

handle_process_control

If set, the program will handle 'suspend', 'resume', and 'quit' messages from the core client. The action depends on the direct_process_action flag.

send_status_msgs

If set, the program will report its CPU time and fraction done to the core client. Set in worker programs.

direct_process_action

If set, the program will respond to quit messages and heartbeat failures by exiting, and will respond to suspend and resume messages by suspending and resuming. Otherwise, these events will result in changes to the BOINC_STATUS structure, which can be polled using boinc_get_status().

all_threads_cpu_time

If set, the CPU of all threads (not just the worker thread) will be counted. For apps that do computation in more than one thread.

Typical main program logic is:

BOINC_OPTIONS options;

options.main_program = true;
...
boinc_init_options(&options)
read main state file
for each remaining worker program:
    APP_INIT_DATA aid;
    boinc_get_init_data(aid);
    aid.fraction_done_start = x
    aid.fraction_done_end = y
    boinc_write_init_data_file(aid)
    run the app
    wait for the app to finish
        poll
    write main state file
    if last app:
        break
    boinc_parse_init_data_file()    // reads CPU time from app_init.xml file
boinc_finish()

where x and y are the appropriate fraction done range limits. Typical worker program logic is:

BOINC_OPTIONS options;

options.main_program = false;
options.send_status_msgs = true;
...
boinc_init_options(&options);
...
do work, calling boinc_fraction_done() with values from 0 to 1,
and boinc_time_to_checkpoint(), occasionally
...
boinc_finish();        // this writes final CPU time to app_init.xml file

If the graphics is handled in a program that runs concurrently with the worker programs, it must also call boinc_init_options(), typically with all options false, then boinc_init_graphics(), and eventually boinc_finish().

If the main program is responsible for reporting application status to the core client, it should periodically call

    boinc_report_app_status(
        double cpu_time,                // CPU time since start of WU
        double checkpoint_cpu_time,     // CPU time at last checkpoint
        double fraction_done
    );