GMPBase.py

Go to the documentation of this file.

from __future__ import print_function
from Gaudi.Configuration import *
from GaudiPython import AppMgr, gbl, setOwnership, PyAlgorithm, SUCCESS, FAILURE, InterfaceCast
from ROOT import TBufferFile, TBuffer
import multiprocessing
from multiprocessing import Process, Queue, JoinableQueue, Event
from multiprocessing import cpu_count, current_process
from multiprocessing.queues import Empty
from pTools import *
import time
import sys
import os
from ROOT import TParallelMergingFile
# This script contains the bases for the Gaudi MultiProcessing (GMP)
# classes

# There are three classes :
#   Reader
#   Worker
#   Writer

# Each class needs to perform communication with the others
# For this, we need a means of communication, which will be based on
# the python multiprocessing package
# This is provided in SPI pytools package
# cmt line : use pytools v1.1 LCG_Interfaces
# The PYTHONPATH env variable may need to be modified, as this might
# still point to 1.0_python2.5

# Each class will need Queues, and a defined method for using these
# queues.
# For example, as long as there is something in the Queue, both ends
# of the queue must be open
# Also, there needs to be proper Termination flags and criteria
# The System should be error proof.

# Constants -------------------------------------------------------------------
NAP = 0.001
MB = 1024.0 * 1024.0
# waits to guard against hanging, in seconds
WAIT_INITIALISE = 60 * 5
WAIT_FIRST_EVENT = 60 * 3
WAIT_SINGLE_EVENT = 60 * 6
WAIT_FINALISE = 60 * 2
STEP_INITIALISE = 10
STEP_EVENT = 2
STEP_FINALISE = 10

# My switch for direct switching on/off Smaps Algorithm in GaudiPython AppMgr
SMAPS = False

# -----------------------------------------------------------------------------

# definitions
# ----------
# used to convert stored histos (in AIDA format) to ROOT format
aida2root = gbl.Gaudi.Utils.Aida2ROOT.aida2root

# used to check which type of histo we are dealing with
# i.e. if currentHisto in aidatypes : pass
aidatypes = (gbl.AIDA.IHistogram, gbl.AIDA.IHistogram1D, gbl.AIDA.IHistogram2D,
             gbl.AIDA.IHistogram3D, gbl.AIDA.IProfile1D, gbl.AIDA.IProfile2D,
             gbl.AIDA.IBaseHistogram)  # extra?

# similar to aidatypes
thtypes = (gbl.TH1D, gbl.TH2D, gbl.TH3D, gbl.TProfile, gbl.TProfile2D)

# Types of OutputStream in Gaudi
WRITERTYPES = {
    'EvtCollectionStream': "tuples",
    'InputCopyStream': "events",
    'OutputStream': "events",
    'RecordStream': "records",
    'RunRecordStream': "records",
    'SequentialOutputStream': "events",
    'TagCollectionStream': "tuples"
}

# =============================================================================


class MiniWriter(object):
    '''
    A class to represent a writer in the GaudiPython configuration
    It can be non-trivial to access the name of the output file; it may be
    specified in the DataSvc, or just on the writer, may be a list, or string
    Also, there are three different types of writer (events, records, tuples)
    so this bootstrap class provides easy access to this info while configuring
    '''

    def __init__(self, writer, wType, config):
        self.w = writer
        self.wType = wType
        # set parameters for directly accessing the correct
        #   part of the configuration, so that later, we can do
        #   config[ key ].Output = modified(output)
        self.key = None
        self.output = None
        self.ItemList = None
        self.OptItemList = None
        #
        self.wName = writer.getName()
        # Now process the Writer, find where the output is named
        self.woutput = None
        self.datasvcName = None
        self.svcOutput = None
        if hasattr(self.w, "Output"):
            self.woutput = self.w.Output
            self.getItemLists(config)
            self.set(self.wName, self.w.Output)
            return
        else:
            # if there is no output file, get it via the datasvc
            # (every writer always has a datasvc property)
            self.datasvcName = self.w.EvtDataSvc
            datasvc = config[self.datasvcName]
            if hasattr(datasvc, "Output"):
                self.getItemLists(config)
                self.set(self.datasvcName, datasvc.Output)
                return

    def getNewName(self, replaceThis, withThis, extra=''):
        # replace one pattern in the output name string
        #  with another, and return the Output name
        # It *might* be in a list, so check for this
        #
        # @param extra : might need to add ROOT flags
        #                i.e.: OPT='RECREATE', or such
        assert replaceThis.__class__.__name__ == 'str'
        assert withThis.__class__.__name__ == 'str'
        old = self.output
        lst = False
        if old.__class__.__name__ == 'list':
            old = self.output[0]
            lst = True
        new = old.replace(replaceThis, withThis)
        new += extra
        if lst:
            return [new]
        else:
            return new

    def getItemLists(self, config):
        # the item list
        if hasattr(self.w, "ItemList"):
            self.ItemList = self.w.ItemList
        else:
            datasvc = config[self.w.EvtDataSvc]
            if hasattr(datasvc, "ItemList"):
                self.ItemList = datasvc.ItemList
        # The option item list; possibly not a valid variable
        if hasattr(self.w, "OptItemList"):
            self.OptItemList = self.w.OptItemList
        else:
            datasvc = config[self.w.EvtDataSvc]
            if hasattr(datasvc, "OptItemList"):
                self.OptItemList = datasvc.OptItemList
        return

    def set(self, key, output):
        self.key = key
        self.output = output
        return

    def __repr__(self):
        s = ""
        line = '-' * 80
        s += (line + '\n')
        s += "Writer         : %s\n" % (self.wName)
        s += "Writer Type    : %s\n" % (self.wType)
        s += "Writer Output  : %s\n" % (self.output)
        s += "DataSvc        : %s\n" % (self.datasvcName)
        s += "DataSvc Output : %s\n" % (self.svcOutput)
        s += '\n'
        s += "Key for config : %s\n" % (self.key)
        s += "Output File    : %s\n" % (self.output)
        s += "ItemList       : %s\n" % (self.ItemList)
        s += "OptItemList    : %s\n" % (self.OptItemList)
        s += (line + '\n')
        return s


# =============================================================================


class CollectHistograms(PyAlgorithm):
    '''
    GaudiPython algorithm used to clean up histos on the Reader and Workers
    Only has a finalize method()
    This retrieves a dictionary of path:histo objects and sends it to the
    writer.  It then waits for a None flag : THIS IS IMPORTANT, as if
    the algorithm returns before ALL histos have been COMPLETELY RECEIVED
    at the writer end, there will be an error.
    '''

    def __init__(self, gmpcomponent):
        PyAlgorithm.__init__(self)
        self._gmpc = gmpcomponent
        self.log = self._gmpc.log
        return None

    def execute(self):
        return SUCCESS

    def finalize(self):
        self.log.info(
            'CollectHistograms Finalise (%s)' % (self._gmpc.nodeType))
        self._gmpc.hDict = self._gmpc.dumpHistograms()
        ks = self._gmpc.hDict.keys()
        self.log.info('%i Objects in Histogram Store' % (len(ks)))

        # crashes occurred due to Memory Error during the sending of hundreds
        # histos on slc5 machines, so instead, break into chunks
        # send 100 at a time
        chunk = 100
        reps = len(ks) / chunk + 1
        for i in range(reps):
            someKeys = ks[i * chunk:(i + 1) * chunk]
            smalld = dict([(key, self._gmpc.hDict[key]) for key in someKeys])
            self._gmpc.hq.put((self._gmpc.nodeID, smalld))
        # "finished" Notification
        self.log.debug('Signalling end of histos to Writer')
        self._gmpc.hq.put('HISTOS_SENT')
        self.log.debug('Waiting on Sync Event')
        self._gmpc.sEvent.wait()
        self.log.debug('Histo Sync Event set, clearing and returning')
        self._gmpc.hvt.clearStore()
        root = gbl.DataObject()
        setOwnership(root, False)
        self._gmpc.hvt.setRoot('/stat', root)
        return SUCCESS


# =============================================================================


class EventCommunicator(object):
    # This class is responsible for communicating Gaudi Events via Queues
    # Events are communicated as TBufferFiles, filled either by the
    # TESSerializer, or the GaudiSvc, "IPCSvc"
    def __init__(self, GMPComponent, qin, qout):
        self._gmpc = GMPComponent
        self.log = self._gmpc.log
        # maximum capacity of Queues
        self.maxsize = 50
        # central variables : Queues
        self.qin = qin
        self.qout = qout

        # flags
        self.allsent = False
        self.allrecv = False

        # statistics storage
        self.nSent = 0  # counter : items sent
        self.nRecv = 0  # counter : items received
        self.sizeSent = 0  # measure : size of events sent ( tbuf.Length() )
        self.sizeRecv = 0  # measure : size of events in   ( tbuf.Length() )
        self.qinTime = 0  # time    : receiving from self.qin
        self.qoutTime = 0  # time    : sending on qout

    def send(self, item):
        # This class manages the sending of a TBufferFile Event to a Queue
        # The actual item to be sent is a tuple : ( evtNumber, TBufferFile )
        assert item.__class__.__name__ == 'tuple'
        startTransmission = time.time()
        self.qout.put(item)
        # allow the background thread to feed the Queue; not 100% guaranteed to
        # finish before next line
        while self.qout._buffer:
            time.sleep(NAP)
        self.qoutTime += time.time() - startTransmission
        self.sizeSent += item[1].Length()
        self.nSent += 1
        return SUCCESS

    def receive(self, timeout=None):
        # Receive items from self.qin
        startWait = time.time()
        try:
            itemIn = self.qin.get(timeout=timeout)
        except Empty:
            return None
        self.qinTime += time.time() - startWait
        self.nRecv += 1
        if itemIn.__class__.__name__ == 'tuple':
            self.sizeRecv += itemIn[1].Length()
        else:
            self.nRecv -= 1
        try:
            self.qin.task_done()
        except:
            self._gmpc.log.warning(
                'TASK_DONE called too often by : %s' % (self._gmpc.nodeType))
        return itemIn

    def finalize(self):
        self.log.info('Finalize Event Communicator : %s %s' %
                      (self._gmpc, self._gmpc.nodeType))
        # Reader sends one flag for each worker
        # Workers send one flag each
        # Writer sends nothing (it's at the end of the chain)
        if self._gmpc.nodeType == 'Reader':
            downstream = self._gmpc.nWorkers
        elif self._gmpc.nodeType == 'Writer':
            downstream = 0
        elif self._gmpc.nodeType == 'Worker':
            downstream = 1

        for i in range(downstream):
            self.qout.put('FINISHED')
        if self._gmpc.nodeType != 'Writer':
            self.qout.join()
        # Now some reporting...
        self.statistics()

    def statistics(self):
        self.log.name = '%s-%i Audit ' % (self._gmpc.nodeType,
                                          self._gmpc.nodeID)
        self.log.info('Items Sent     : %i' % (self.nSent))
        self.log.info('Items Received : %i' % (self.nRecv))
        self.log.info('Data  Sent     : %i' % (self.sizeSent))
        self.log.info('Data  Received : %i' % (self.sizeRecv))
        self.log.info('Q-out Time     : %5.2f' % (self.qoutTime))
        self.log.info('Q-in  Time     : %5.2f' % (self.qinTime))


# =============================================================================


class TESSerializer(object):
    def __init__(self, gaudiTESSerializer, evtDataSvc, nodeType, nodeID, log):
        self.T = gaudiTESSerializer
        self.evt = evtDataSvc
        self.buffersIn = []
        self.buffersOut = []
        self.nIn = 0
        self.nOut = 0
        self.tDump = 0.0
        self.tLoad = 0.0
        # logging
        self.nodeType = nodeType
        self.nodeID = nodeID
        self.log = log

    def Load(self, tbuf):
        root = gbl.DataObject()
        setOwnership(root, False)
        self.evt.setRoot('/Event', root)
        t = time.time()
        self.T.loadBuffer(tbuf)
        self.tLoad += (time.time() - t)
        self.nIn += 1
        self.buffersIn.append(tbuf.Length())

    def Dump(self):
        t = time.time()
        tb = TBufferFile(TBuffer.kWrite)
        self.T.dumpBuffer(tb)
        self.tDump += (time.time() - t)
        self.nOut += 1
        self.buffersOut.append(tb.Length())
        return tb

    def Report(self):
        evIn = "Events Loaded    : %i" % (self.nIn)
        evOut = "Events Dumped    : %i" % (self.nOut)
        din = sum(self.buffersIn)
        dataIn = "Data Loaded      : %i" % (din)
        dataInMb = "Data Loaded (MB) : %5.2f Mb" % (din / MB)
        if self.nIn:
            avgIn = "Avg Buf Loaded   : %5.2f Mb"\
                % (din / (self.nIn * MB))
            maxIn = "Max Buf Loaded   : %5.2f Mb"\
                % (max(self.buffersIn) / MB)
        else:
            avgIn = "Avg Buf Loaded   : N/A"
            maxIn = "Max Buf Loaded   : N/A"
        dout = sum(self.buffersOut)
        dataOut = "Data Dumped      : %i" % (dout)
        dataOutMb = "Data Dumped (MB) : %5.2f Mb" % (dout / MB)
        if self.nOut:
            avgOut = "Avg Buf Dumped   : %5.2f Mb"\
                % (din / (self.nOut * MB))
            maxOut = "Max Buf Dumped   : %5.2f Mb"\
                % (max(self.buffersOut) / MB)
        else:
            avgOut = "Avg Buf Dumped   : N/A"
            maxOut = "Max Buf Dumped   : N/A"
        dumpTime = "Total Dump Time  : %5.2f" % (self.tDump)
        loadTime = "Total Load Time  : %5.2f" % (self.tLoad)

        lines = evIn,\
            evOut,\
            dataIn,\
            dataInMb,\
            avgIn,\
            maxIn,\
            dataOut,\
            dataOutMb,\
            avgOut,\
            maxOut,\
            dumpTime,\
            loadTime
        self.log.name = "%s-%i TESSerializer" % (self.nodeType, self.nodeID)
        for line in lines:
            self.log.info(line)
        self.log.name = "%s-%i" % (self.nodeType, self.nodeID)


# =============================================================================


class GMPComponent(object):
    # This class will be the template for Reader, Worker and Writer
    # containing all common components
    # nodeId will be a numerical identifier for the node
    # -1 for reader
    # -2 for writer
    # 0,...,nWorkers-1 for the Workers
    def __init__(self, nodeType, nodeID, queues, events, params, subworkers):
        # declare a Gaudi MultiProcessing Node
        # the nodeType is going to be one of Reader, Worker, Writer
        # qPair is going to be a tuple of ( qin, qout )
        # for sending and receiving
        # if nodeType is "Writer", it will be a list of qPairs,
        # as there's one queue-in from each Worker
        #
        # params is a tuple of (nWorkers, config, log)

        self.nodeType = nodeType
        current_process().name = nodeType

        # Synchronisation Event() objects for keeping track of the system
        self.initEvent, eventLoopSyncer, self.finalEvent = events
        self.eventLoopSyncer, self.lastEvent = eventLoopSyncer  # unpack tuple

        # necessary for knowledge of the system
        self.nWorkers, self.sEvent, self.config, self.log = params
        self.subworkers = subworkers
        self.nodeID = nodeID
        self.msgFormat = self.config['MessageSvc'].Format

        # describe the state of the node by the current Event Number
        self.currentEvent = None

        # Unpack the Queues : (events, histos, filerecords)
        self.queues = queues
        self.num = 0

        ks = self.config.keys()
        self.app = None
        list = ["Brunel", "DaVinci", "Boole", "Gauss"]
        for k in list:
            if k in ks:
                self.app = k

    def Start(self):
        # define the separate process
        qPair, histq, fq = self.queues

        # Set up the Queue Mechanisms ( Event Communicators )
        if self.nodeType == 'Reader' or self.nodeType == 'Worker':
            # Reader or Worker Node
            qin, qout = qPair
            self.evcom = EventCommunicator(self, qin, qout)
        else:
            # Writer : many queues in, no queue out
            assert self.nodeType == 'Writer'
            self.evcoms = []
            qsin = qPair[0]
            for q in qsin:
                ec = EventCommunicator(self, q, None)
                self.evcoms.append(ec)
        # Histogram Queue
        self.hq = histq
        # FileRecords Queue
        self.fq = fq

        # Universal Counters (available to all nodes)
        # Use sensibly!!!
        self.nIn = 0
        self.nOut = 0

        # Status Flag (possibly remove later)
        self.stat = SUCCESS

        # Set logger name
        self.log.name = '%s-%i' % (self.nodeType, self.nodeID)

        # Heuristic variables
        # time for init, run, final, firstEventTime, totalTime
        self.iTime = 0.0
        self.rTime = 0.0
        self.fTime = 0.0
        self.firstEvTime = 0.0
        self.tTime = 0.0

        self.proc = Process(target=self.Engine)
        # Fork and start the separate process
        self.proc.start()

    def Engine(self):
        # This will be the method carried out by the Node
        # Different for all
        pass

    def processConfiguration(self):
        # Different for all ; customize Configuration for multicore
        pass

    def SetupGaudiPython(self):
        # This method will initialize the GaudiPython Tools
        # such as the AppMgr and so on
        self.a = AppMgr()
        if SMAPS:
            from AlgSmapShot import SmapShot
            smapsLog = self.nodeType + '-' + str(self.nodeID) + '.smp'
            ss = SmapShot(logname=smapsLog)
            self.a.addAlgorithm(ss)
        self.evt = self.a.evtsvc()
        self.hvt = self.a.histsvc()
        self.fsr = self.a.filerecordsvc()
        self.inc = self.a.service('IncidentSvc', 'IIncidentSvc')
        self.pers = self.a.service('EventPersistencySvc', 'IAddressCreator')
        self.ts = gbl.GaudiMP.TESSerializer(self.evt._idp, self.pers)
        self.TS = TESSerializer(self.ts, self.evt, self.nodeType, self.nodeID,
                                self.log)
        return SUCCESS

    def StartGaudiPython(self):
        self.a.initialize()
        self.a.start()
        return SUCCESS

    def LoadTES(self, tbufferfile):
        root = gbl.DataObject()
        setOwnership(root, False)
        self.evt.setRoot('/Event', root)
        self.ts.loadBuffer(tbufferfile)

    def getEventNumber(self):
        if self.app != 'Gauss':
            # Using getList or getHistoNames can result in the EventSelector
            # re-initialising connection to RootDBase, which costs a lot of
            # time... try to build a set of Header paths??

            # First Attempt : Unpacked Event Data
            lst = ['/Event/Gen/Header', '/Event/Rec/Header']
            for l in lst:
                path = l
                try:
                    n = self.evt[path].evtNumber()

                    return n
                except:
                    # No evt number at this path
                    continue

            # second attepmt : try DAQ/RawEvent data
            # The Evt Number is in bank type 16, bank 0, data pt 4
            try:
                n = self.evt['/Event/DAQ/RawEvent'].banks(16)[0].data()[4]

                return n
            except:
                pass

            # Default Action
            if self.nIn > 0 or self.nOut > 0:
                pass
            else:
                self.log.warning('Could not determine Event Number')
            return -1
        else:
            if self.nodeID == -1:
                self.num = self.num + 1

                return self.num

    def IdentifyWriters(self):
        #
        # Identify Writers in the Configuration
        #
        d = {}
        keys = ["events", "records", "tuples", "histos"]
        for k in keys:
            d[k] = []

        # Identify Writers and Classify
        wkeys = WRITERTYPES.keys()
        for v in self.config.values():
            if v.__class__.__name__ in wkeys:
                writerType = WRITERTYPES[v.__class__.__name__]
                d[writerType].append(MiniWriter(v, writerType, self.config))
                if self.nodeID == 0:
                    self.log.info('Writer Found : %s' % (v.name()))

        # Now Check for the Histogram Service
        if 'HistogramPersistencySvc' in self.config.keys():
            hfile = self.config['HistogramPersistencySvc'].getProp(
                'OutputFile')
            d["histos"].append(hfile)
        return d

    def dumpHistograms(self):
        '''
        Method used by the GaudiPython algorithm CollectHistos
        to obtain a dictionary of form { path : object }
        representing the Histogram Store
        '''
        nlist = self.hvt.getHistoNames()
        histDict = {}
        objects = 0
        histos = 0
        if nlist:
            for n in nlist:
                o = self.hvt[n]
                if type(o) in aidatypes:
                    o = aida2root(o)
                    histos += 1
                else:
                    objects += 1
                histDict[n] = o
        else:
            print('WARNING : no histograms to recover?')
        return histDict

    def Initialize(self):
        start = time.time()
        self.processConfiguration()
        self.SetupGaudiPython()
        self.initEvent.set()
        self.StartGaudiPython()

        if self.app == 'Gauss':

            tool = self.a.tool("ToolSvc.EvtCounter")
            self.cntr = InterfaceCast(gbl.IEventCounter)(tool.getInterface())
        else:
            self.cntr = None

        self.iTime = time.time() - start

    def Finalize(self):
        start = time.time()
        self.a.stop()
        self.a.finalize()
        self.log.info('%s-%i Finalized' % (self.nodeType, self.nodeID))
        self.finalEvent.set()
        self.fTime = time.time() - start

    def Report(self):
        self.log.name = "%s-%i Audit" % (self.nodeType, self.nodeID)
        allTime = "Alive Time     : %5.2f" % (self.tTime)
        initTime = "Init Time      : %5.2f" % (self.iTime)
        frstTime = "1st Event Time : %5.2f" % (self.firstEvTime)
        runTime = "Run Time       : %5.2f" % (self.rTime)
        finTime = "Finalise Time  : %5.2f" % (self.fTime)
        tup = (allTime, initTime, frstTime, runTime, finTime)
        for t in tup:
            self.log.info(t)
        self.log.name = "%s-%i" % (self.nodeType, self.nodeID)
        # and report from the TESSerializer
        self.TS.Report()


# =============================================================================


class Reader(GMPComponent):
    def __init__(self, queues, events, params, subworkers):
        GMPComponent.__init__(self, 'Reader', -1, queues, events, params,
                              subworkers)

    def processConfiguration(self):
        # Reader :
        #   No algorithms
        #   No output
        #   No histos
        self.config['ApplicationMgr'].TopAlg = []
        self.config['ApplicationMgr'].OutStream = []
        if "HistogramPersistencySvc" in self.config.keys():
            self.config['HistogramPersistencySvc'].OutputFile = ''
        self.config['MessageSvc'].Format = '%-13s ' % '[Reader]' + \
            self.msgFormat
        self.evtMax = self.config['ApplicationMgr'].EvtMax

    def DumpEvent(self):
        tb = TBufferFile(TBuffer.kWrite)
        # print '----Reader dumping Buffer!!!'
        self.ts.dumpBuffer(tb)
        # print '\tBuffer Dumped, size : %i'%( tb.Length() )
        return tb

    def DoFirstEvent(self):
        # Do First Event ------------------------------------------------------
        # Check Termination Criteria
        startFirst = time.time()
        self.log.info('Reader : First Event')
        if self.nOut == self.evtMax:
            self.log.info('evtMax( %i ) reached' % (self.evtMax))
            self.lastEvent.set()
            return SUCCESS
        else:
            # Continue to read, dump and send event
            self.a.run(1)
            if not bool(self.evt['/Event']):
                self.log.warning('No More Events! (So Far : %i)' % (self.nOut))
                self.lastEvent.set()
                return SUCCESS
            else:
                # Popluate TESSerializer list and send Event
                if self.app == "Gauss":
                    lst = self.evt.getHistoNames()
                else:
                    try:
                        lst = self.evt.getList()
                        if self.app == "DaVinci":
                            daqnode = self.evt.retrieveObject(
                                '/Event/DAQ').registry()
                            setOwnership(daqnode, False)
                            self.evt.getList(daqnode, lst,
                                             daqnode.address().par())
                    except:
                        self.log.critical('Reader could not acquire TES List!')
                        self.lastEvent.set()
                        return FAILURE
                self.log.info('Reader : TES List : %i items' % (len(lst)))
                for l in lst:
                    self.ts.addItem(l)
                self.currentEvent = self.getEventNumber()
                tb = self.TS.Dump()
                self.log.info('First Event Sent')
                self.evcom.send((self.currentEvent, tb))
                self.nOut += 1
                self.eventLoopSyncer.set()
                self.evt.clearStore()
                self.firstEvTime = time.time() - startFirst
                return SUCCESS

    def Engine(self):
        # rename process
        import os
        import ctypes
        libc = ctypes.CDLL('libc.so.6')
        name = str(self.nodeType) + str(self.nodeID) + '\0'
        libc.prctl(15, name, 0, 0, 0)

        startEngine = time.time()
        self.log.name = 'Reader'
        self.log.info('Reader Process starting')

        self.Initialize()

        # add the Histogram Collection Algorithm
        self.a.addAlgorithm(CollectHistograms(self))

        self.log.info('Reader Beginning Distribution')
        sc = self.DoFirstEvent()
        if sc.isSuccess():
            self.log.info('Reader First Event OK')
        else:
            self.log.critical('Reader Failed on First Event')
            self.stat = FAILURE

        # Do All Others -------------------------------------------------------
        while True:
            # Check Termination Criteria
            if self.nOut == self.evtMax:
                self.log.info('evtMax( %i ) reached' % (self.evtMax))
                break
            # Check Health
            if not self.stat.isSuccess():
                self.log.critical('Reader is Damaged!')
                break
            # Continue to read, dump and send event
            t = time.time()
            self.a.run(1)
            self.rTime += (time.time() - t)
            if not bool(self.evt['/Event']):
                self.log.warning('No More Events! (So Far : %i)' % (self.nOut))
                break
            self.currentEvent = self.getEventNumber()
            tb = self.TS.Dump()
            self.evcom.send((self.currentEvent, tb))
            # clean up
            self.nOut += 1
            self.eventLoopSyncer.set()
            self.evt.clearStore()
        self.log.info('Setting <Last> Event')
        self.lastEvent.set()

        # Finalize
        self.log.info('Reader : Event Distribution complete.')
        self.evcom.finalize()
        self.Finalize()
        self.tTime = time.time() - startEngine
        self.Report()


# =============================================================================


class Subworker(GMPComponent):
    def __init__(self, workerID, queues, events, params, subworkers):
        GMPComponent.__init__(self, 'Worker', workerID, queues, events, params,
                              subworkers)
        # Identify the writer streams
        self.writerDict = self.IdentifyWriters()
        # Identify the accept/veto checks for each event
        self.acceptAlgs, self.requireAlgs, self.vetoAlgs = self.getCheckAlgs()
        self.log.info("Subworker-%i Created OK" % (self.nodeID))
        self.eventOutput = True

    def Engine(self):
        # rename process
        import os
        import ctypes
        libc = ctypes.CDLL('libc.so.6')
        name = str(self.nodeType) + str(self.nodeID) + '\0'
        libc.prctl(15, name, 0, 0, 0)

        self.initEvent.set()
        startEngine = time.time()
        msg = self.a.service('MessageSvc')
        msg.Format = '%-13s ' % ('[' + self.log.name + ']') + self.msgFormat

        self.log.name = "Worker-%i" % (self.nodeID)
        self.log.info("Subworker %i starting Engine" % (self.nodeID))
        self.filerecordsAgent = FileRecordsAgent(self)

        # populate the TESSerializer itemlist
        self.log.info(
            'EVT WRITERS ON WORKER : %i' % (len(self.writerDict['events'])))

        nEventWriters = len(self.writerDict["events"])
        self.a.addAlgorithm(CollectHistograms(self))

        # Begin processing
        Go = True
        while Go:
            packet = self.evcom.receive()
            if packet:
                pass
            else:
                continue
            if packet == 'FINISHED':
                break
            evtNumber, tbin = packet  # unpack
            if self.cntr != None:

                self.cntr.setEventCounter(evtNumber)

            self.nIn += 1
            self.TS.Load(tbin)

            t = time.time()
            sc = self.a.executeEvent()
            if self.nIn == 1:
                self.firstEvTime = time.time() - t
            else:
                self.rTime += (time.time() - t)
            if sc.isSuccess():
                pass
            else:
                self.log.name = "Worker-%i" % (self.nodeID)
                self.log.warning('Did not Execute Event')
                self.evt.clearStore()
                continue
            if self.isEventPassed():
                pass
            else:
                self.log.name = "Worker-%i" % (self.nodeID)
                self.log.warning('Event did not pass : %i' % (evtNumber))
                self.evt.clearStore()
                continue
            if self.eventOutput:
                # It may be the case of generating Event Tags; hence
                #   no event output
                self.currentEvent = self.getEventNumber()
                tb = self.TS.Dump()
                self.evcom.send((self.currentEvent, tb))
                self.nOut += 1
            self.inc.fireIncident(gbl.Incident('Subworker', 'EndEvent'))
            self.eventLoopSyncer.set()
            self.evt.clearStore()
        self.log.name = "Worker-%i" % (self.nodeID)
        self.log.info('Setting <Last> Event %s' % (self.nodeID))
        self.lastEvent.set()

        self.evcom.finalize()
        # Now send the FileRecords and stop/finalize the appMgr
        self.filerecordsAgent.SendFileRecords()
        self.tTime = time.time() - startEngine
        self.Finalize()
        self.Report()
        # self.finalEvent.set()

    def SetServices(self, a, evt, hvt, fsr, inc, pers, ts, cntr):
        self.a = a
        self.evt = evt
        self.hvt = hvt
        self.fsr = fsr
        #self.inc = inc
        self.inc = self.a.service('IncidentSvc', 'IIncidentSvc')
        self.pers = pers
        self.ts = ts
        self.cntr = cntr
        self.TS = TESSerializer(self.ts, self.evt, self.nodeType, self.nodeID,
                                self.log)

    def getCheckAlgs(self):
        '''
        For some output writers, a check is performed to see if the event has
        executed certain algorithms.
        These reside in the AcceptAlgs property for those writers
        '''
        acc = []
        req = []
        vet = []
        for m in self.writerDict["events"]:
            if hasattr(m.w, 'AcceptAlgs'):
                acc += m.w.AcceptAlgs
            if hasattr(m.w, 'RequireAlgs'):
                req += m.w.RequireAlgs
            if hasattr(m.w, 'VetoAlgs'):
                vet += m.w.VetoAlgs
        return (acc, req, vet)

    def checkExecutedPassed(self, algName):
        if self.a.algorithm(algName)._ialg.isExecuted()\
                and self.a.algorithm(algName)._ialg.filterPassed():
            return True
        else:
            return False

    def isEventPassed(self):
        '''
        Check the algorithm status for an event.
        Depending on output writer settings, the event
          may be declined based on various criteria.
        This is a transcript of the check that occurs in GaudiSvc::OutputStream
        '''
        passed = False

        self.log.debug('self.acceptAlgs is %s' % (str(self.acceptAlgs)))
        if self.acceptAlgs:
            for name in self.acceptAlgs:
                if self.checkExecutedPassed(name):
                    passed = True
                    break
        else:
            passed = True

        self.log.debug('self.requireAlgs is %s' % (str(self.requireAlgs)))
        for name in self.requireAlgs:
            if self.checkExecutedPassed(name):
                pass
            else:
                self.log.info('Evt declined (requireAlgs) : %s' % (name))
                passed = False

        self.log.debug('self.vetoAlgs is %s' % (str(self.vetoAlgs)))
        for name in self.vetoAlgs:
            if self.checkExecutedPassed(name):
                pass
            else:
                self.log.info('Evt declined : (vetoAlgs) : %s' % (name))
                passed = False
        return passed


# =============================================================================


class Worker(GMPComponent):
    def __init__(self, workerID, queues, events, params, subworkers):
        GMPComponent.__init__(self, 'Worker', workerID, queues, events, params,
                              subworkers)
        # Identify the writer streams
        self.writerDict = self.IdentifyWriters()
        # Identify the accept/veto checks for each event
        self.acceptAlgs, self.requireAlgs, self.vetoAlgs = self.getCheckAlgs()
        self.log.name = "Worker-%i" % (self.nodeID)
        self.log.info("Worker-%i Created OK" % (self.nodeID))
        self.eventOutput = True

    def processConfiguration(self):

        # Worker :
        #   No input
        #   No output
        #   No Histos
        self.config['EventSelector'].Input = []
        self.config['ApplicationMgr'].OutStream = []
        if "HistogramPersistencySvc" in self.config.keys():
            self.config['HistogramPersistencySvc'].OutputFile = ''
        formatHead = '[Worker-%i]' % (self.nodeID)
        self.config['MessageSvc'].Format = '%-13s ' % formatHead + \
            self.msgFormat

        for key, lst in self.writerDict.iteritems():
            self.log.info('Writer Type : %s\t : %i' % (key, len(lst)))

        for m in self.writerDict["tuples"]:
            # rename Tuple output file with an appendix
            # based on worker id, for merging later
            newName = m.getNewName('.', '.w%i.' % (self.nodeID))
            self.config[m.key].Output = newName

        # Suppress INFO Output for all but Worker-0
        # if self.nodeID == 0 :
        #    pass
        # else                :
        #    self.config[ 'MessageSvc' ].OutputLevel = ERROR

        if self.app == "Gauss":
            try:
                if "ToolSvc.EvtCounter" not in self.config:
                    from Configurables import EvtCounter
                    counter = EvtCounter()
                else:
                    counter = self.config["ToolSvc.EvtCounter"]
                counter.UseIncident = False
            except:
                # ignore errors when trying to change the configuration of the EvtCounter
                self.log.warning('Cannot configure EvtCounter')

    def Engine(self):

        # rename process
        import os
        import ctypes
        libc = ctypes.CDLL('libc.so.6')
        name = str(self.nodeType) + str(self.nodeID) + '\0'
        libc.prctl(15, name, 0, 0, 0)

        startEngine = time.time()
        self.log.info("Worker %i starting Engine" % (self.nodeID))
        self.Initialize()
        self.filerecordsAgent = FileRecordsAgent(self)

        # populate the TESSerializer itemlist
        self.log.info(
            'EVT WRITERS ON WORKER : %i' % (len(self.writerDict['events'])))

        nEventWriters = len(self.writerDict["events"])
        if nEventWriters:
            itemList = set()
            optItemList = set()
            for m in self.writerDict["events"]:
                for item in m.ItemList:
                    hsh = item.find('#')
                    if hsh != -1:
                        item = item[:hsh]
                    itemList.add(item)
                for item in m.OptItemList:
                    hsh = item.find('#')
                    if hsh != -1:
                        item = item[:hsh]
                    optItemList.add(item)
            # If an item is mandatory and optional, keep it only in the optional list
            itemList -= optItemList
            for item in sorted(itemList):
                self.log.info(' adding ItemList Item to ts : %s' % (item))
                self.ts.addItem(item)
            for item in sorted(optItemList):
                self.log.info(' adding Optional Item to ts : %s' % (item))
                self.ts.addOptItem(item)
        else:
            self.log.info('There is no Event Output for this app')
            self.eventOutput = False

        # Begin processing
        Go = True
        while Go:
            packet = self.evcom.receive()
            if packet:
                pass
            else:
                continue
            if packet == 'FINISHED':
                break
            evtNumber, tbin = packet  # unpack
            if self.cntr != None:
                self.cntr.setEventCounter(evtNumber)

            # subworkers are forked before the first event is processed
            if self.nIn == 0:
                self.log.info("Fork new subworkers")

                # Fork subworkers and share services
                for k in self.subworkers:
                    k.SetServices(self.a, self.evt, self.hvt, self.fsr,
                                  self.inc, self.pers, self.ts, self.cntr)
                    k.Start()
                    self.a.addAlgorithm(CollectHistograms(self))
            self.nIn += 1
            self.TS.Load(tbin)

            t = time.time()
            sc = self.a.executeEvent()
            if self.nIn == 1:
                self.firstEvTime = time.time() - t
            else:
                self.rTime += (time.time() - t)
            if sc.isSuccess():
                pass
            else:
                self.log.warning('Did not Execute Event')
                self.evt.clearStore()
                continue
            if self.isEventPassed():
                pass
            else:
                self.log.warning('Event did not pass : %i' % (evtNumber))
                self.evt.clearStore()
                continue
            if self.eventOutput:
                # It may be the case of generating Event Tags; hence
                #   no event output
                self.currentEvent = self.getEventNumber()
                tb = self.TS.Dump()
                self.evcom.send((self.currentEvent, tb))
                self.nOut += 1
            self.inc.fireIncident(gbl.Incident('Worker', 'EndEvent'))
            self.eventLoopSyncer.set()
            self.evt.clearStore()
        self.log.info('Setting <Last> Event')
        self.lastEvent.set()

        self.evcom.finalize()
        self.log.info('Worker-%i Finished Processing Events' % (self.nodeID))
        # Now send the FileRecords and stop/finalize the appMgr
        self.filerecordsAgent.SendFileRecords()
        self.Finalize()
        self.tTime = time.time() - startEngine
        self.Report()

        for k in self.subworkers:
            self.log.info('Join subworkers')
            k.proc.join()

    def getCheckAlgs(self):
        '''
        For some output writers, a check is performed to see if the event has
        executed certain algorithms.
        These reside in the AcceptAlgs property for those writers
        '''
        acc = []
        req = []
        vet = []
        for m in self.writerDict["events"]:
            if hasattr(m.w, 'AcceptAlgs'):
                acc += m.w.AcceptAlgs
            if hasattr(m.w, 'RequireAlgs'):
                req += m.w.RequireAlgs
            if hasattr(m.w, 'VetoAlgs'):
                vet += m.w.VetoAlgs
        return (acc, req, vet)

    def checkExecutedPassed(self, algName):
        if self.a.algorithm(algName)._ialg.isExecuted()\
                and self.a.algorithm(algName)._ialg.filterPassed():
            return True
        else:
            return False

    def isEventPassed(self):
        '''
        Check the algorithm status for an event.
        Depending on output writer settings, the event
          may be declined based on various criteria.
        This is a transcript of the check that occurs in GaudiSvc::OutputStream
        '''
        passed = False

        self.log.debug('self.acceptAlgs is %s' % (str(self.acceptAlgs)))
        if self.acceptAlgs:
            for name in self.acceptAlgs:
                if self.checkExecutedPassed(name):
                    passed = True
                    break
        else:
            passed = True

        self.log.debug('self.requireAlgs is %s' % (str(self.requireAlgs)))
        for name in self.requireAlgs:
            if self.checkExecutedPassed(name):
                pass
            else:
                self.log.info('Evt declined (requireAlgs) : %s' % (name))
                passed = False

        self.log.debug('self.vetoAlgs is %s' % (str(self.vetoAlgs)))
        for name in self.vetoAlgs:
            if self.checkExecutedPassed(name):
                pass
            else:
                self.log.info('Evt declined : (vetoAlgs) : %s' % (name))
                passed = False
        return passed


# =============================================================================


class Writer(GMPComponent):
    def __init__(self, queues, events, params, subworkers):
        GMPComponent.__init__(self, 'Writer', -2, queues, events, params,
                              subworkers)
        # Identify the writer streams
        self.writerDict = self.IdentifyWriters()
        # This keeps track of workers as they finish
        self.status = [False] * self.nWorkers
        self.log.name = "Writer--2"

    def processConfiguration(self):
        # Writer :
        #   No input
        #   No Algs
        self.config['ApplicationMgr'].TopAlg = []
        self.config['EventSelector'].Input = []

        self.config['MessageSvc'].Format = '%-13s ' % '[Writer]' + \
            self.msgFormat

        # Now process the output writers
        for key, lst in self.writerDict.iteritems():
            self.log.info('Writer Type : %s\t : %i' % (key, len(lst)))

        # Modify the name of the output file to reflect that it came
        #  from a parallel processing
        #
        # Event Writers
        for m in self.writerDict["events"]:
            self.log.debug('Processing Event Writer : %s' % (m))
            newName = m.getNewName('.', '.p%i.' % self.nWorkers)
            self.config[m.key].Output = newName

        # Now, if there are no event writers, the FileRecords file
        #   will fail to open, as it only opens an UPDATE version
        #   of the existing Event Output File
        # So, if there are no event writers, edit the string of the
        #   FileRecord Writer

        # FileRecords Writers
        for m in self.writerDict["records"]:
            self.log.debug('Processing FileRecords Writer: %s' % (m))
            newName = m.getNewName(
                '.', '.p%i.' % self.nWorkers, extra=" OPT='RECREATE'")
            self.config[m.key].Output = newName

        # same for histos
        hs = "HistogramPersistencySvc"
        n = None
        if hs in self.config.keys():
            n = self.config[hs].OutputFile
        if n:
            newName = self.config[hs].OutputFile.replace(
                '.', '.p%i.' % (self.nWorkers))
            self.config[hs].OutputFile = newName

    def Engine(self):
        # rename process
        import os
        import ctypes
        libc = ctypes.CDLL('libc.so.6')
        name = str(self.nodeType) + str(self.nodeID) + '\0'
        libc.prctl(15, name, 0, 0, 0)

        startEngine = time.time()
        self.Initialize()
        self.histoAgent = HistoAgent(self)
        self.filerecordsAgent = FileRecordsAgent(self)

        # Begin processing
        Go = True
        current = -1
        stopCriteria = self.nWorkers
        while Go:
            current = (current + 1) % self.nWorkers
            packet = self.evcoms[current].receive(timeout=0.01)
            if packet == None:
                continue
            if packet == 'FINISHED':
                self.log.info(
                    'Writer got FINISHED flag : Worker %i' % (current))

                self.status[current] = True
                if all(self.status):
                    self.log.info('FINISHED recd from all workers, break loop')
                    break
                continue
            # otherwise, continue as normal
            self.nIn += 1  # update central count (maybe needed by FSR store)
            evtNumber, tbin = packet  # unpack
            self.TS.Load(tbin)
            t = time.time()
            self.a.executeEvent()
            self.rTime += (time.time() - t)
            self.currentEvent = self.getEventNumber()
            self.evt.clearStore()
            self.eventLoopSyncer.set()
        self.log.name = "Writer--2"
        self.log.info('Setting <Last> Event')
        self.lastEvent.set()

        # finalisation steps
        [e.finalize() for e in self.evcoms]
        # Now do Histograms
        sc = self.histoAgent.Receive()
        sc = self.histoAgent.RebuildHistoStore()
        if sc.isSuccess():
            self.log.info('Histo Store rebuilt ok')
        else:
            self.log.warning('Histo Store Error in Rebuild')

        # Now do FileRecords
        sc = self.filerecordsAgent.Receive()
        self.filerecordsAgent.Rebuild()
        self.Finalize()
        #self.rTime = time.time()-startEngine
        self.Report()


# =============================================================================

# =============================================================================


class Coord(object):
    def __init__(self, nWorkers, config, log):

        self.log = log
        self.config = config
        # set up Logging
        self.log.name = 'GaudiPython-Parallel-Logger'
        self.log.info('GaudiPython Parallel Process Co-ordinator beginning')

        if nWorkers == -1:
            # The user has requested all available cpus in the machine
            self.nWorkers = cpu_count()
        else:
            self.nWorkers = nWorkers

        self.qs = self.SetupQueues()  # a dictionary of queues (for Events)
        self.hq = JoinableQueue()  # for Histogram data
        self.fq = JoinableQueue()  # for FileRecords data

        # Make a Syncer for Initalise, Run, and Finalise
        self.sInit = Syncer(
            self.nWorkers,
            self.log,
            limit=WAIT_INITIALISE,
            step=STEP_INITIALISE)
        self.sRun = Syncer(
            self.nWorkers,
            self.log,
            manyEvents=True,
            limit=WAIT_SINGLE_EVENT,
            step=STEP_EVENT,
            firstEvent=WAIT_FIRST_EVENT)
        self.sFin = Syncer(
            self.nWorkers, self.log, limit=WAIT_FINALISE, step=STEP_FINALISE)
        # and one final one for Histogram Transfer
        self.histSyncEvent = Event()

        # params are common to al subprocesses
        params = (self.nWorkers, self.histSyncEvent, self.config, self.log)

        self.subworkers = []
        # Declare SubProcesses!
        for i in range(1, self.nWorkers):
            sub = Subworker(i, self.getQueues(i), self.getSyncEvents(i),
                            params, self.subworkers)
            self.subworkers.append(sub)
        self.reader = Reader(
            self.getQueues(-1), self.getSyncEvents(-1), params,
            self.subworkers)
        self.workers = []
        wk = Worker(0, self.getQueues(0), self.getSyncEvents(0), params,
                    self.subworkers)
        self.writer = Writer(
            self.getQueues(-2), self.getSyncEvents(-2), params,
            self.subworkers)

        self.system = []
        self.system.append(self.writer)
        self.system.append(wk)
        self.system.append(self.reader)

    def getSyncEvents(self, nodeID):
        init = self.sInit.d[nodeID].event
        run = (self.sRun.d[nodeID].event, self.sRun.d[nodeID].lastEvent)
        fin = self.sFin.d[nodeID].event
        return (init, run, fin)

    def getQueues(self, nodeID):
        eventQ = self.qs[nodeID]
        histQ = self.hq
        fsrQ = self.fq
        return (eventQ, histQ, fsrQ)

    def Go(self):

        # Initialise
        self.log.name = 'GaudiPython-Parallel-Logger'
        self.log.info('INITIALISING SYSTEM')

        # Start reader, writer and main worker
        for p in self.system:
            p.Start()

        sc = self.sInit.syncAll(step="Initialise")
        if sc == SUCCESS:
            pass
        else:
            self.Terminate()
            return FAILURE

        # Run
        self.log.name = 'GaudiPython-Parallel-Logger'
        self.log.info('RUNNING SYSTEM')
        sc = self.sRun.syncAll(step="Run")
        if sc == SUCCESS:
            pass
        else:
            self.Terminate()
            return FAILURE

        # Finalise
        self.log.name = 'GaudiPython-Parallel-Logger'
        self.log.info('FINALISING SYSTEM')
        sc = self.sFin.syncAll(step="Finalise")
        if sc == SUCCESS:
            pass
        else:
            self.Terminate()
            return FAILURE

        # if we've got this far, finally report SUCCESS
        self.log.info("Cleanly join all Processes")
        self.Stop()
        self.log.info("Report Total Success to Main.py")
        return SUCCESS

    def Terminate(self):
        # Brutally kill sub-processes
        children = multiprocessing.active_children()
        for i in children:
            i.terminate()

        # self.writer.proc.terminate()
        #[ w.proc.terminate() for w in self.workers]
        # self.reader.proc.terminate()

    def Stop(self):
        # procs should be joined in reverse order to launch
        self.system.reverse()
        for s in self.system:
            s.proc.join()
        return SUCCESS

    def SetupQueues(self):
        # This method will set up the network of Queues needed
        # N Queues = nWorkers + 1
        # Each Worker has a Queue in, and a Queue out
        # Reader has Queue out only
        # Writer has nWorkers Queues in

        # one queue from Reader-Workers
        rwk = JoinableQueue()
        # one queue from each worker to writer
        workersWriter = [JoinableQueue() for i in range(self.nWorkers)]
        d = {}
        d[-1] = (None, rwk)  # Reader
        d[-2] = (workersWriter, None)  # Writer
        for i in range(self.nWorkers):
            d[i] = (rwk, workersWriter[i])
        return d


# ============================= EOF ===========================================