ManySmallAlgs.py

Go to the documentation of this file.

#!/usr/bin/env gaudirun.py
 
"""
A totally synthetic test of scheduler performance, with a large number of CF nodes and algorithms, but each algorithm does no work.
 
In the flat mode (nested=False) 1000 algorithms are added directly to the base CF node as a baseline measurement.
 
In nested mode the algorithms are grouped into sets of under a parent CF node.
Those nodes are then given 2 different CF node parents (20 in total) which include a prescale algorithm.
 
In sequential mode (sequence=True) the prescale algorithms will randomly deactivate CF nodes at a frequency given by filterPass
"""
from Configurables import AlgResourcePool, AvalancheSchedulerSvc
from Configurables import Gaudi__Examples__Prescaler as Prescaler
from Configurables import Gaudi__Sequencer, HiveSlimEventLoopMgr, HiveWhiteBoard
from Gaudi.Configuration import *
 
# metaconfig -------------------------------------------------------------------
# It's confortable to collect the relevant parameters at the top of the optionfile
evtslots = 5
evtMax = 1000
cardinality = 10
threads = 5
nested = True
sequence = True
filterPass = 50.0
 
# -------------------------------------------------------------------------------
 
# The configuration of the whiteboard ------------------------------------------
# It is useful to call it EventDataSvc to replace the usual data service with
# the whiteboard transparently.
 
whiteboard = HiveWhiteBoard("EventDataSvc", EventSlots=evtslots, OutputLevel=FATAL)
 
# -------------------------------------------------------------------------------
 
# Event Loop Manager -----------------------------------------------------------
# It's called slim since it has less functionalities overall than the good-old
# event loop manager. Here we just set its outputlevel to DEBUG.
 
slimeventloopmgr = HiveSlimEventLoopMgr(
    SchedulerName="AvalancheSchedulerSvc", OutputLevel=INFO
)
 
# -------------------------------------------------------------------------------
 
# ForwardScheduler -------------------------------------------------------------
# We just decide how many algorithms in flight we want to have and how many
# threads in the pool. The default value is -1, which is for TBB equivalent
# to take over the whole machine.
 
scheduler = AvalancheSchedulerSvc(
    ThreadPoolSize=threads,
    ShowControlFlow=False,
    ShowDataDependencies=False,
    OutputLevel=INFO,
)
 
# -------------------------------------------------------------------------------
 
# Algo Resource Pool -----------------------------------------------------------
# Nothing special here, we just set the debug level.
AlgResourcePool(OutputLevel=FATAL)
 
# -------------------------------------------------------------------------------
 
# Flat structure
allAlgs = []
for i in range(1000):
    alg = Prescaler("alg" + str(i))
    alg.OutputLevel = FATAL
    alg.Cardinality = cardinality
    alg.PercentPass = 100.0
    allAlgs.append(alg)
 
baseSeq = Gaudi__Sequencer(
    "baseSeq",
    Members=allAlgs,
    Sequential=sequence,
    ModeOR=True,
    ShortCircuit=False,
    OutputLevel=FATAL,
)
 
# Nested structure
allSeqs = []
allSeqSeqs = []
if nested:
    for i in range(100):
        seqAlgs = allAlgs[10 * i : 10 * (i + 1)]
        seq = Gaudi__Sequencer(
            "seq" + str(i),
            Members=seqAlgs,
            Sequential=sequence,
            ModeOR=False,
            ShortCircuit=True,
            OutputLevel=FATAL,
        )
        allSeqs.append(seq)
 
    for i in range(10):
        seqSeqs = allSeqs[10 * i : 10 * (i + 1)]
 
        filterAlg = Prescaler("filterAlg" + str(i))
        filterAlg.OutputLevel = FATAL
        filterAlg.Cardinality = cardinality
        filterAlg.PercentPass = filterPass
 
        seq = Gaudi__Sequencer(
            "seqSeq" + str(i),
            Members=[filterAlg] + seqSeqs,
            Sequential=sequence,
            ModeOR=False,
            ShortCircuit=True,
            OutputLevel=FATAL,
        )
        allSeqSeqs.append(seq)
 
    for i in range(9):
        seqSeqs = allSeqs[5 + (10 * i) : 15 + (10 * i)]
 
        filterAlg = Prescaler("filterAlg" + str(i + 10))
        filterAlg.OutputLevel = FATAL
        filterAlg.Cardinality = cardinality
        filterAlg.PercentPass = filterPass
 
        seq = Gaudi__Sequencer(
            "seqSeq" + str(i + 10),
            Members=[filterAlg] + seqSeqs,
            Sequential=sequence,
            ModeOR=False,
            ShortCircuit=True,
            OutputLevel=FATAL,
        )
        allSeqSeqs.append(seq)
 
    baseSeq.Members = allSeqSeqs
 
# Application Manager ----------------------------------------------------------
# We put everything together and change the type of message service
 
ApplicationMgr(
    EvtMax=evtMax,
    EvtSel="NONE",
    ExtSvc=[whiteboard],
    EventLoop=slimeventloopmgr,
    TopAlg=[baseSeq],
    MessageSvcType="InertMessageSvc",
)
 
# -------------------------------------------------------------------------------