-
Notifications
You must be signed in to change notification settings - Fork 449
JobIn
A BOINC job has two parts:
- A workunit describing the computation to be performed.
- One or more results, each of which describes an instance of a computation, either unstarted, in progress, or completed. The BOINC client software refers to results as "tasks".
These entities are stored in the 'workunit' and 'result' database tables respectively.
BOINC provides a utility program and C function for creating jobs.
A workunit has several attributes. These are specified when the workunit is created; they are mandatory except where noted.
A text string, unique across all workunits in the project. You can guarantee uniqueness by embedding the PID of the creating process, a sequence number, and a Unix timestamp in the name.
Which application will perform the computation. A workunit is associated with an application, not with a particular version or range of versions. If the input data format changes in a way that is incompatible with older versions, you must either a) release new versions for all supported platforms, or b) create a new application. Such incompatibilities can often be avoided by using XML data format.
A list of the input files: their names, and the names by which the application refers to them.
Typically these file are downloaded from a data server.
However, if the <generate_locally/> element is present, the file is generated on the client
(typically by an earlier instance of the same application).
Applications should use file locking to prevent two jobs from generating the file at the same time.
(optional) An integer; can be used to operate (cancel, change priority etc.) on groups of workunits.
It's important that you supply accurate values for these parameters. See estimating job sizes for tips on how to do so.
An estimate of the number of floating-point operations required to complete the job, used to estimate how long the job will take on a given host.
An upper bound on the number of floating-point operations required to complete the job. If this bound is exceeded, the job will be aborted.
An estimate of job's largest working set size. The job will only be sent to hosts with at least this much available RAM.
A bound on the maximum disk space used by the job, including all input, temporary, and output files. The job will only be sent to hosts with at least this much available disk space. If this bound is exceeded, the job will be aborted.
If nonzero, this job will be sent only to hosts with at least this much download bandwidth. Use for jobs with very large input files.
An upper bound on the time (in seconds) between sending a result to a client and receiving a reply. The scheduler won't issue a result if the estimated completion time exceeds this. If the client doesn't respond within this interval, the server 'gives up' on the result and generates a new result, to be assigned to another client. Set this to several times the average execution time of a workunit on a typical PC. If you set it too low, BOINC may not be able to send some results, and the corresponding workunit will be flagged with an error. If you set it too high, there may a corresponding delay in getting results back.
The minimum size of a 'quorum'. The validator is run when there are this many successful results. If a strict majority agree, they are considered correct. Set this to two or more if you want redundant computing.
How many results to create initially. This must be at least min_quorum. It may be more, to reflect the ratio of result loss, or to get a quorum more quickly.
If the number of client error results exceeds this, the work unit is declared to have an error; no further results are issued, and the assimilator is triggered. This safeguards against workunits that cause the application to crash.
If the total number of results for this workunit would exceed this, the workunit is declared to be in error. This safeguards against workunits that are never reported (e.g. because they crash the core client).
If the number of success results for this workunit exceeds this, and a consensus has not been reached, the workunit is declared to be in error. This safeguards against workunits that produce nondeterministic results.
(optional) Higher-priority work is dispatched first
A workunit can experience any of several error conditions:
The BOINC scheduler was unable to send the workunit to a large number (~100) of hosts, probably because its resource requirements (disk, memory, CPU) were too large for the hosts, or because no application version was available for the hosts' platforms. In this case BOINC 'gives up' on the workunit.
Too many results with error conditions (upload/download problem, client crashes) have been returned for this work unit.
Too many successful results have been returned without consensus. This indicates that the application may be nondeterministic.
Too many total results have been sent for this workunit.
If any of these conditions holds, BOINC doesn't dispatch more instances of the workunit.
The main attribute of a result is list of the names of the output files, and the names by which the application refers to them.
A result has a server_state whose values include:
| Inactive (not ready to dispatch) | |
| Unsent (ready to send to a client, but not sent) | |
| In progress (sent, not done) | |
| Done successfully | |
| Timed out | |
| Done with error | |
| Not needed (work unit was finalized before this result was sent) |
Additional attributes are defined after the result is completed:
| host | The host that executed the computation. |
| exit status | The exit status (0 if success). |
| CPU time | The CPU time that was used. |
| output file info | The sizes and checksums of its output files. |
| stderr | The stderr output of the computation. |
| received time | The time when the result was received. |