class TaskJuggler::BatchProcessor
The BatchProcessor
class can be used to run code blocks of the program as a separate process. Mulitple pieces of code can be submitted to be executed in parallel. The number of CPU cores to use is limited at object creation time. The submitted jobs will be queued and scheduled to the given number of CPUs. The usage model is simple. Create an BatchProcessor
object. Use BatchProcessor#queue
to submit all the jobs and then use BatchProcessor#wait
to wait for completion and to process the results.
Public Class Methods
Create a BatchProcessor
object. maxCpuCores
limits the number of simultaneously spawned processes.
# File lib/taskjuggler/BatchProcessor.rb, line 69 def initialize(maxCpuCores) @maxCpuCores = maxCpuCores # Jobs submitted by calling queue() are put in the @toRunQueue. The # launcher Thread will pick them up and fork them off into another # process. @toRunQueue = [ ] # A hash that maps the JobInfo objects of running jobs by their PID. @runningJobs = { } # A list of jobs that wait to complete their writing. @spoolingJobs = [ ] # The wait() method will then clean the @toDropQueue, executes the post # processing block and removes all JobInfo related objects. @toDropQueue = [] # A semaphore to guard accesses to @runningJobs, @spoolingJobs and # following shared data structures. @lock = Monitor.new # We count the submitted and completed jobs. The @jobsIn counter also # doubles as a unique job ID. @jobsIn = @jobsOut = 0 # An Array that holds all the IO objects to receive data from. @pipes = [] # A hash that maps IO objects to JobInfo objects @pipeToJob = {} # This global flag is set to true to signal the threads to terminate. @terminate = false # Sleep time of the threads when no data is pending. This value must be # large enough to allow for a context switch between the sending # (forked-off) process and this process. If it's too large, throughput # will suffer. @timeout = 0.02 Thread.abort_on_exception = true end
Public Instance Methods
Add a new job the job queue. tag
is some data that the caller can use to identify the job upon completion. block
is a Ruby code block to be executed in a separate process.
# File lib/taskjuggler/BatchProcessor.rb, line 108 def queue(tag = nil, &block) # Create a new JobInfo object for the job and push it to the @toRunQueue. @lock.synchronize do raise 'You cannot call queue() while wait() is running!' if @jobsOut > 0 # If this is the first queued job for this run, we have to start the # helper threads. if @jobsIn == 0 # The JobInfo objects in the @toRunQueue are processed by the # launcher thread. It forkes off processes to execute the code # block associated with the JobInfo. @launcher = Thread.new { launcher } # The receiver thread waits for terminated child processes and picks # up the results. @receiver = Thread.new { receiver } # The grabber thread collects $stdout and $stderr data from each # child process and stores them in the corresponding JobInfo. @grabber = Thread.new { grabber } end # To track a job through the queues, we use a JobInfo object to hold # all data associated with a job. job = JobInfo.new(@jobsIn, block, tag) # Increase job counter @jobsIn += 1 # Push the job to the toRunQueue. @toRunQueue.push(job) end end
Wait for all jobs to complete. The code block will get the JobInfo
objects for each job to pick up the results.
# File lib/taskjuggler/BatchProcessor.rb, line 141 def wait # Don't wait if there are no jobs. return if @jobsIn == 0 # When we have received as many jobs in the @toDropQueue than we have # started then we're done. while @lock.synchronize { @jobsOut < @jobsIn } job = nil @lock.synchronize do if !@toDropQueue.empty? && (job = @toDropQueue.pop) # Call the post-processing block that was passed to wait() with # the JobInfo object as argument. @jobsOut += 1 yield(job) end end unless job sleep(@timeout) end end # Signal threads to stop @terminate = true # Wait for treads to finish @launcher.join @receiver.join @grabber.join # Reset some variables so we can reuse the object for further job runs. @jobsIn = @jobsOut = 0 @terminate = false # Make sure all data structures are empty and clean. check end