wiki:RemoteJobs

Version 59 (modified by davea, 7 years ago) (diff)

--

RPCs for remote job submission

This document describes APIs for remotely submitting, monitoring, and controlling jobs on a BOINC server. The APIs supports the submission of batches of jobs, which may contain a single job or many thousands of jobs. Currently, the API has two restrictions:

  • All jobs in a batch must use the same application.
  • There can be no dependencies between jobs.

At the bottom level, the interface consists of Web RPCs. BOINC provides client-side bindings in PHP, C++, and Python. These bindings differ slightly; they expose different details of the Web RPCs.

Interfaces for staging input files and fetching output files are described separately.

There are various options for managing input files. If you use Job-based file management, which maintains batch/file associations, the order of operations is:

  • Create a batch (initially empty); returns the batch ID.
  • Stage input files, passing the batch ID
  • Submit jobs, passing the batch ID

If you manage input files a different way, then you create the batch and submit jobs in a single API call.

Once you have submitted the batch, you can

  • Monitor the batch with query_batches(), query_batch(), or query_job().
  • Abort the batch (if you see errors, or if enough jobs have been finished) using abort_jobs() or abort_batch().
  • download output files.
  • Retire the batch using retire_batch(). This tells the server to clean up the files and job records associated with the batch, and to mark the batch as "retired"; retired batches are normally not shown in the web interface.

PHP interface

The following functions are provided in the PHP file submit.inc, which is independent of other BOINC PHP code. The file html/user/submit_test.php has code to exercise and test these functions.

boinc_submit_batch()

Submits a batch.

Arguments: a "request object" whose fields include

  • project: the project URL
  • authenticator: the user's authenticator
  • app_name: the name of the application for which jobs are being submitted
  • batch_name: a symbolic name for the batch. Must be unique. If omitted, a name of the form "batch_unixtime" will be used.
  • workunit_template_file: an optional input template file name.
  • result_template_file: an optional output template file name.
  • app_version_num: if present, pins the jobs to a particular app version number.
  • jobs: an array of job descriptors, each of which contains
    • rsc_fpops_est: an estimate of the FLOPs used by the job
    • command_line: command-line arguments to the application
    • wu_template: optional; the input template to use for this job, as an XML string.
    • result_template: optional; the output template to use for this job, as an XML string.
    • input_files: an array of input file descriptors, each of which contains
      • mode: "local", "semilocal", "local_staged", "inline", or "remote" (see below).
      • source: meaning depends on mode:
        • local: path on the BOINC server
        • semilocal: the file's URL
        • local_staged: physical name
        • inline: the file's contents
      • For "remote" mode, instead of "source" you must specify:
        • url: the file's URL
        • nbytes: file size
        • md5: the file's MD5 checksum

Result: a 2-element array containing

  • The batch ID
  • An error message (null if success)

Input files can be supplied in any of the following ways:

  • local: the file is on the BOINC server and is not staged. It's specified by its full path.
  • local_staged: the filed has been staged on the BOINC server. It's specified by its physical name.
  • semilocal: the file is on a data server that's accessible to the BOINC server but not necessarily to the outside world. The file is specified by its URL. It will be downloaded by the BOINC server during job submission, and served to clients from the BOINC server.
  • inline: the file is included in the job submission request XML message. It will be served to clients from BOINC server.
  • remote: the file is on a data server other than the BOINC server, and will be served to clients from that data server. It's specified by the URL, the file size, and the file MD5.

The following mode has been proposed but is not implemented yet:

  • sandbox: the file is in the user's sandbox, and is specified by its name in the sandbox.

The following example submits a 10-job batch:

$req = new StdClass;
$req->project = "http://foo.bar.edu/test/";
$req->authenticator = "xxx";
$req->app_name = "uppercase";
$req->jobs = array();

$f = new StdClass;
$f->mode = "local_staged";
$f->source = "filename.dat";

$job = new StdClass;
$job->input_files = array($f);

for ($i=10; $i<20; $i++) {
    $job->rsc_fpops_est = $i*1e9;
    $job->command_line = "--t $i";
    $req->jobs[] = $job;
}

list($batch_id, $errmsg) = boinc_submit_batch($req);
if ($errmsg) {
    echo "Error: $errmsg\n";
} else {
    echo "Batch ID: $batch_id\n";
}

Note: this interfaces is inconsistent; it lets you do some things but not others. Let me know if you need additions.

boinc_estimate_batch()

Returns an estimate of the elapsed time required to complete a batch.

Arguments: same as boinc_submit_batch() (only relevant fields need to be populated).

Return value: a 2-element array containing

  • The elapsed time estimate, in seconds
  • An error message (null if success)

boinc_query_batches()

Returns a list of this user's batches, both in progress and complete.

Argument: a request object with elements

  • project and authenticator: as above.
  • get_cpu_time (optional): if nonzero, get CPU time of each batch

Result: a 2-element array. The first element is an array of batch descriptor objects, each with the following fields:

  • id: batch ID
  • state: values are
    • 1: in progress
    • 2: completed (all jobs either succeeded or had fatal errors)
    • 3: aborted
    • 4: retired
  • name: the batch name
  • app_name: the application name
  • create_time: when the batch was submitted
  • est_completion_time: current estimate of completion time
  • njobs: the number of jobs in the batch
  • fraction_done: the fraction of the batch that has been completed (0..1)
  • nerror_jobs: the number of jobs that had fatal errors
  • completion_time: when the batch was completed
  • credit_estimate: BOINC's initial estimate of the credit that would be granted to complete the batch, including replication
  • credit_canonical: the actual credit granted to canonical instances
  • credit_total: the actual credit granted to all instances

boinc_query_batch()

Gets batch details.

Argument: a request object with elements

  • project and authenticator: as above
  • batch_id: specifies a batch.
  • get_cpu_time (optional): if nonzero, get CPU time of batch. This includes all job instances, and doesn't include GPU time, so it may not be meaningful.
  • get_job_details (optional): if nonzero, return job details (see below).

Result: a 2-element array. The first element is a batch descriptor object as described above, with an additional element:

  • jobs: an array of job descriptor objects, each one containing
    • id: the database ID of the job's workunit record
    • canonical_instance_id: if the job has a canonical instance, its database ID

If get_job_details was set, the job descriptors also contain:

  • status: "queued", "in_progress", "error", or "done".
  • cpu_time: if done, the CPU time of canonical instance.
  • exit_status: if error, the exit status of one of the error instances.

The order of job descriptors matches their order in the batch submission.

boinc_query_job()

Gets job details.

Argument: a request object with elements:

  • project and authenticator: as above
  • job_id: specifies a job.

Result: a 2-element array. The first element is a job descriptor object with the following fields:

  • instances: an array of job instance descriptors, each containing:
    • name: the instance's name
    • id: the ID of the corresponding result record
    • state: a string describing the instance's state (unsent, in progress, complete, etc.)
    • outfile: if the instance is over, a list of output file descriptors, each containing
      • size: file size in bytes

boinc_abort_batch()

Argument: a request object with elements

  • project and authenticator: as above,
  • batch_id: specifies a batch.

Result: an error message, null if successful

boinc_retire_batch()

Delete server storage (files, DB records) associated with a batch.

Argument: a request object with elements

  • project and authenticator: as above,
  • batch_id: specifies a batch.

Result: an error message, null if successful

C++ interface

A C++ interface to the following functions is available in lib/remote_submit.cpp. Include lib/remote_submit.h.

All functions return zero on success, else an error code as defined in lib/error_numbers.h

create_batch()

Create a batch - a set of jobs, initially empty.

int create_batch(
    const char* project_url,
    const char* authenticator,
    const char* batch_name,
    const char* app_name,
    double expire_time,
    int &batch_id,
    string& error_msg
);
project_url
the project URL
authenticator
the authenticator of the submitting user
batch_name
a name for the batch. Must be unique over all batches.
app_name
the name of an application on the BOINC server
expire_time
if nonzero, the Unix time when the batch should be aborted and removed from the server, whether or not it's completed.
batch_id
(out) the batch's database ID
error_msg
(out) an error message if the operation failed

estimate_batch()

Get an estimate of the makespan of a (potential) batch.

int estimate_batch(
    const char* project_url,
    const char* authenticator,
    char app_name[256],
    vector<JOB> jobs,
    double est_makespan,
    string& error_msg
);
jobs
description of jobs; same as for submit_jobs() (see below).
est_makespan
the estimated makespan (time to completion).

submit_jobs()

Submit a set of jobs; place them in an existing batch, and make them runnable.

int submit_jobs(
    const char* project_url,
    const char* authenticator,
    char app_name[256],
    int batch_id,
    vector<JOB> jobs,
    string& error_msg
);

struct JOB {
    char job_name[256];
    string cmdline_args;
    vector<INFILE> infiles;
};

struct INFILE {
    int mode;
       // FILE_MODE_LOCAL_STAGED: file is already on BOINC server, and staged
       // FILE_MODE_REMOTE: file is on a different server

    // the following if LOCAL_STAGED
    char physical_name[256];

    // the following if REMOTE
    char url[256];
    double nbytes;
    char md5[256];
};
batch_id
ID of a previously created batch

For each job:

job_name
must be unique over all jobs
cmdline_args
command-line arguments
infiles
list of input files

For each input file:

physical_name
BOINC's physical name for the file. The file must already be staged.

query_batches()

Query the status of this user's batches.

int query_batches(
    const char* project_url,
    const char* authenticator,
    vector<BATCH_STATUS>& batches,
    string& error_msg
);

struct BATCH_STATUS {
    int id;
    char name[256];             // name of batch
    char app_name[256];
    int state;                  // see lib/common_defs.h
    int njobs;                  // how many jobs in batch
    int nerror_jobs;            // how many jobs errored out
    double fraction_done;       // how much of batch is done (0..1)
    double create_time;         // when batch was created
    double expire_time;         // when it will expire
    double est_completion_time;     // estimated completion time
    double completion_time;     // if completed, actual completion time
    double credit_estimate;     // original estimate for credit
    double credit_canonical;    // if completed, granted credit

    int parse(XML_PARSER&);
    void print();
};

query_batch()

Return the detailed status of jobs in a given batch (can specify by either ID or name).

extern int query_batch(
    const char* project_url,
    const char* authenticator,
    int batch_id,
    const char batch_name[256],
    vector<JOB_STATE>& jobs,
    string& error_msg
);

struct JOB_STATE {
    int id;
    char name[256];
    int canonical_instance_id;      // it job completed successfully,
                                    // the ID of the canonical instance
    int n_outfiles;                 // number of output files
};

abort_jobs()

Abort a set of jobs.

extern int abort_jobs(
    const char* project_url,
    const char* authenticator,
    vector<string> &job_names,
    string& error_msg
);

query_completed_job()

Query a completed job.

extern int query_completed_job(
    const char* project_url,
    const char* authenticator,
    const char* job_name,
    COMPLETED_JOB_DESC&,
    string& error_msg
);

struct COMPLETED_JOB_DESC {
    int canonical_resultid;
    int error_mask;
    int error_resultid;
    int exit_status;
    double elapsed_time;
    double cpu_time;
    string stderr_out;
};
canonical_resultid
database ID of the "canonical" instance of the job.
error_mask
a bitmask of error conditions (see db/boinc_db_types.h)
error_resultid
the database ID of a failed instance, if one exists
exit_status
exit status of failed instance
elapsed_time
run time of canonical instance
cpu_time
CPU time of canonical instance
stderr_out
stderr output of canonical or failed instance

retire_batch()

"Retire" a batch. The server is then allowed to delete the batch's input and output files, and its database records.

extern int retire_batch(
    const char* project_url,
    const char* authenticator,
    const char* batch_name,
    string& error_msg
);

set_expire_time()

Change the expiration time of a batch.

extern int set_expire_time(
    const char* project_url,
    const char* authenticator,
    const char* batch_name,
    double expire_time,
    string& error_msg
);

ping_server()

Ping the project's server; return zero if the server is up.

extern int ping_server(
    const char* project_url,
    string& error_msg
);

query_batch_set()

Return the status of the jobs in a given set of batches. This is used by the Condor interface; it's probably not useful outside of that.

extern int query_batch_set(
    const char* project_url,
    const char* authenticator,
    vector<string> &batch_names,
    QUERY_BATCH_REPLY& reply,
    string& error_msg
);

struct JOB_STATUS {
    string job_name;
    string status;           // DONE, ERROR, or IN_PROGRESS
};

struct QUERY_BATCH_SET_REPLY {
    vector<int> batch_sizes;    // how many jobs in each of the queried batches
    vector<JOB_STATUS> jobs;   // the jobs, sequentially
};

Python binding

The file tools/submit_api.py contains a Python binding of some of above RPCs. For examples of its use, see tools/submit_api_test.py.

To build a description of a batch of jobs, use the BATCH_DESC, JOB_DESC, and FILE_DESC classes:

def make_batch():
    file = FILE_DESC()
    file.mode = 'remote'
    file.url = 'http://isaac.ssl.berkeley.edu/validate_logic.txt'
    file.md5 = "eec5a142cea5202c9ab2e4575a8aaaa7"
    file.nbytes = 4250

    job = JOB_DESC()
    job.files = [file]

    batch = BATCH_DESC()
    batch.project = 'http://isaac.ssl.berkeley.edu/test/'
    batch.authenticator = get_auth()
    batch.app_name = "uppercase"
    batch.batch_name = "blah"
    # or batch_id = n, if you previously created an empty batch.
    batch.jobs = []

    for i in range(3):
        job.rsc_fpops_est = i*1e9
        job.command_line = '-i %s' %(i)
        batch.jobs.append(copy.copy(job))

    return batch

You can then pass this to either estimate_batch() or submit_batch():

    from submit_api import *
    batch = make_batch()
    r = estimate_batch(batch)
    if r.tag == 'error':
        print 'error: ', r.find('error_msg').text
        return
    print 'estimated time: ', r.text, ' seconds'

The return value of all the API functions is an EntityTree representation of the XML returned by the RPC.

Other requests use a REQUEST object. For example, to query the status of batch 271:

    req = REQUEST()
    req.project = 'http://isaac.ssl.berkeley.edu/test/'
    req.authenticator = get_auth()
    req.batch_id = 271
    req.get_cpu_time = True
    r = query_batch(req)
    if r[0].tag == 'error':
        print 'error: ', r[0].find('error_msg').text
        return
    print 'njobs: ', r.find('njobs').text
    print 'fraction done: ', r.find('fraction_done').text
    print 'total CPU time: ', r.find('total_cpu_time').text
    # ... various other fields
    print 'jobs:'
    for job in r.findall('job'):
        print '   id: ', job.find('id').text
        print '      n_outfiles: ', job.find('n_outfiles').text
        # ... various other fields

Possible attributes of FILE_DESC:

  • mode
  • url
  • nbytes
  • md5
  • source

Possible attributes of JOB_DESC:

  • rsc_fpops
  • command_line
  • wu_template
  • result_template
  • files (list of FILE_DESC)

Possible attributes of BATCH_DESC:

  • project (URL of project)
  • authenticator (submitter's account key)
  • app_name
  • batch_id
  • batch_name
  • app_version_num (for job pinning)
  • jobs (list of JOB_DESC)

Available functions are listed below. Each function takes a request object whose attributes include at least project and authenticator.

abort_batch(req)
req attributes: batch_id
abort_jobs(req)
req attributes: jobs (list of job names)
create_batch(req)
req attributes: app_name, batch_name, expire_time
estimate_batch(req)
req is a BATCH_DESC
query_batch(req)
req attributes: batch_id, get_cpu_time
query_batches(req)
req attributes_ get_cpu_time
query_completed_job(req)
req attributes: job_name
query_job(req)
req attributes: job_id
get_output_file(req)
req attributes: instance_name, file_num
get_output_files(req)
req attributes: batch_id
retire_batch(req)
req attributes: batch_id

HTTP/XML interface

At a lower level, the APIs are accessed by sending a POST request, using HTTP or HTTPS, to PROJECT_URL/submit_rpc_handler.php. The inputs and outputs of each function are XML documents. The format of the request and reply XML documents can be inferred from user/submit_rpc_handler.php.

Example web interface

An example of a web interface for job submission and control, based on this API, can be found here: http://boinc.berkeley.edu/trac/browser/trunk/boinc/html/user/submit_example.php

This example is functional and it shows how to use the API. However, you will have to modify it for your particular applications and web site.