# **************************************************************************
# *
# * Authors: J.M. De la Rosa Trevin (jmdelarosa@cnb.csic.es)
# *
# * Unidad de Bioinformatica of Centro Nacional de Biotecnologia, CSIC
# *
# * This program is free software; you can redistribute it and/or modify
# * it under the terms of the GNU General Public License as published by
# * the Free Software Foundation; either version 3 of the License, or
# * (at your option) any later version.
# *
# * This program is distributed in the hope that it will be useful,
# * but WITHOUT ANY WARRANTY; without even the implied warranty of
# * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# * GNU General Public License for more details.
# *
# * You should have received a copy of the GNU General Public License
# * along with this program; if not, write to the Free Software
# * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
# * 02111-1307 USA
# *
# * All comments concerning this program package may be sent to the
# * e-mail address 'scipion@cnb.csic.es'
# *
# **************************************************************************
"""
This module have the classes for execution of protocol steps.
The basic one will run steps, one by one, after completion.
There is one based on threads to execute steps in parallel
using different threads and the last one with MPI processes.
"""
import logging
logger = logging.getLogger(__name__)
import time
import datetime
import threading
import os
import pyworkflow.utils.process as process
from pyworkflow.utils.path import getParentFolder, removeExt
from . import constants as cts
from .launch import _submit, UNKNOWN_JOBID, _checkJobStatus
[docs]class StepExecutor:
""" Run a list of Protocol steps. """
def __init__(self, hostConfig, **kwargs):
self.hostConfig = hostConfig
self.gpuList = kwargs.get(cts.GPU_LIST, None)
self.protocol = None
[docs] def getGpuList(self):
""" Return the GPU list assigned to current thread. """
return self.gpuList
[docs] def setProtocol(self, protocol):
""" Set protocol to append active jobs to its jobIds. """
self.protocol = protocol
[docs] def runJob(self, log, programName, params,
numberOfMpi=1, numberOfThreads=1,
env=None, cwd=None, executable=None):
""" This function is a wrapper around runJob,
providing the host configuration.
"""
process.runJob(log, programName, params,
numberOfMpi, numberOfThreads,
self.hostConfig,
env=env, cwd=cwd, gpuList=self.getGpuList(), executable=executable)
def _getRunnable(self, steps, n=1):
""" Return the n steps that are 'new' and all its
dependencies have been finished, or None if none ready.
"""
rs = [] # return a list of runnable steps
for s in steps:
if (s.getStatus() == cts.STATUS_NEW and
all(steps[i-1].isFinished() for i in s._prerequisites)):
if self._isStepRunnable(s):
rs.append(s)
if len(rs) == n:
break
return rs
def _isStepRunnable(self, step):
""" Should be implemented by inherited classes to test extra conditions """
return True
def _arePending(self, steps):
""" Return True if there are pending steps (either running or waiting)
that can be done and thus enable other steps to be executed.
"""
return any(s.isRunning() or s.isWaiting() for s in steps)
[docs] def runSteps(self, steps,
stepStartedCallback,
stepFinishedCallback,
stepsCheckCallback,
stepsCheckSecs=3):
# Even if this will run the steps in a single thread
# let's follow a similar approach than the parallel one
# In this way we can take into account the steps graph
# dependency and also the case when using streaming
delta = datetime.timedelta(seconds=stepsCheckSecs)
lastCheck = datetime.datetime.now()
while True:
# Get a step to run, if there is any
runnableSteps = self._getRunnable(steps)
if runnableSteps:
step = runnableSteps[0]
# We found a step to work in, so let's start a new
# thread to do the job and book it.
step.setRunning()
stepStartedCallback(step)
step.run()
doContinue = stepFinishedCallback(step)
if not doContinue:
break
elif self._arePending(steps):
# We have not found any runnable step, but still there
# there are some running or waiting for dependencies
# So, let's wait a bit to check if something changes
time.sleep(0.5)
else:
# No steps to run, neither running or waiting
# So, we are done, either failed or finished :)
break
now = datetime.datetime.now()
if now - lastCheck > delta:
stepsCheckCallback()
lastCheck = now
stepsCheckCallback() # one last check to finalize stuff
[docs]class StepThread(threading.Thread):
""" Thread to run Steps in parallel. """
def __init__(self, step, lock):
threading.Thread.__init__(self)
self.thId = step.getObjId()
self.step = step
self.lock = lock
[docs] def run(self):
error = None
try:
self.step._run() # not self.step.run() , to avoid race conditions
except Exception as e:
error = str(e)
logger.error("Couldn't run the code in a thread." , exc_info=e)
finally:
with self.lock:
if error is None:
self.step.setFinished()
else:
self.step.setFailed(error)
[docs]class ThreadStepExecutor(StepExecutor):
""" Run steps in parallel using threads. """
def __init__(self, hostConfig, nThreads, **kwargs):
StepExecutor.__init__(self, hostConfig, **kwargs)
self.numberOfProcs = nThreads
# If the gpuList was specified, we need to distribute GPUs among
# all the threads
self.gpuDict = {}
self._assignGPUperNode()
def _assignGPUperNode(self):
# If we have GPUs
if self.gpuList:
nThreads = self.numberOfProcs
# Nodes: each concurrent steps
nodes = range(1, nThreads+1)
# Number of GPUs
nGpu = len(self.gpuList)
# If more GPUs than threads
if nGpu > nThreads:
# Get the ratio: 2 GPUs per thread? 3 GPUs per thread?
# 3 GPU and 2 threads is rounded to 1 (flooring)
step = int(nGpu / nThreads)
spare = nGpu % nThreads
fromPos = 0
# For each node(concurrent thread)
for node in nodes:
# Store the GPUS per thread:
# GPUs: 0 1 2
# Threads 2 (step 1)
# Node 0 : GPU 0 1
# Node 1 : GPU 2
extraGpu = 1 if spare>0 else 0
toPos = fromPos + step +extraGpu
gpusForNode = list(self.gpuList[fromPos:toPos])
newGpusForNode = self.cleanVoidGPUs(gpusForNode)
if len(newGpusForNode) == 0:
logger.info("Gpu slot cancelled: all were null Gpus -> %s" % gpusForNode)
else:
logger.info("GPUs %s assigned to node %s" % (newGpusForNode, node))
self.gpuDict[-node] = newGpusForNode
fromPos = toPos
spare-=1
else:
# Expand gpuList repeating until reach nThreads items
if nThreads > nGpu:
logger.warning("GPUs are no longer extended. If you want all GPUs to match threads repeat as many "
"GPUs as threads.")
# newList = self.gpuList * (int(nThreads / nGpu) + 1)
# self.gpuList = newList[:nThreads]
for index, gpu in enumerate(self.gpuList):
if gpu == cts.VOID_GPU:
logger.info("Void GPU (%s) found in the list. Skipping the slot." % cts.VOID_GPU)
else:
logger.info("GPU slot for gpu %s." % gpu)
# Any negative number in the key means a free gpu slot. can't be 0!
self.gpuDict[-index-1] = [gpu]
[docs] def cleanVoidGPUs(self, gpuList):
newGPUList=[]
for gpuid in gpuList:
if gpuid == cts.VOID_GPU:
logger.info("Void GPU detected in %s" % gpuList)
else:
newGPUList.append(gpuid)
return newGPUList
[docs] def getGpuList(self):
""" Return the GPU list assigned to current thread
or empty list if not using GPUs. """
# If the node id has assigned gpus?
nodeId = threading.current_thread().thId
if nodeId in self.gpuDict:
gpus = self.gpuDict.get(nodeId)
logger.info("Reusing GPUs (%s) slot for %s" % (gpus, nodeId))
return gpus
else:
gpus = self.getFreeGpuSlot(nodeId)
if gpus is None:
logger.warning("Step on node %s is requesting GPUs but there isn't any available. Review configuration of threads/GPUs. Returning and empty list." % nodeId)
return []
else:
return gpus
[docs] def getFreeGpuSlot(self, stepId=None):
""" Returns a free gpu slot available or None. If node is passed it also reserves it for that node
:param node: node to make the reserve of Gpus
"""
for node in self.gpuDict.keys():
# This is a free node. Book it
if node < 0:
gpus = self.gpuDict[node]
if stepId is not None:
self.gpuDict.pop(node)
self.gpuDict[stepId] = gpus
logger.info("GPUs %s assigned to step %s" % (gpus, stepId))
else:
logger.info("Free gpu slot found at %s" % node)
return gpus
return None
[docs] def freeGpusSlot(self, node):
gpus = self.gpuDict.get(node, None)
# Some nodes/threads do not use gpus so may not be booked and not in the dictionary
if gpus is not None:
self.gpuDict.pop(node)
self.gpuDict[-node] = gpus
logger.info("GPUs %s freed from step %s" % (gpus, node))
else:
logger.debug("step id %s not found in GPU slots" % node)
def _isStepRunnable(self, step):
""" Overwrite this method to check GPUs availability"""
if self.gpuList and step.needsGPU() and self.getFreeGpuSlot(step.getObjId()) is None:
logger.info("Can't run step %s. Needs gpus and there are no free gpu slots" % step)
return False
return True
[docs] def runSteps(self, steps,
stepStartedCallback,
stepFinishedCallback,
stepsCheckCallback,
stepsCheckSecs=5):
"""
Creates threads and synchronize the steps execution.
:param steps: list of steps to run
:param stepStartedCallback: callback to be called before starting any step
:param stepFinishedCallback: callback to be run after all steps are done
:param stepsCheckCallback: callback to check if there are new steps to add (streaming)
:param stepsCheckSecs: seconds between stepsCheckCallback calls
"""
delta = datetime.timedelta(seconds=stepsCheckSecs)
lastCheck = datetime.datetime.now()
sharedLock = threading.Lock()
runningSteps = {} # currently running step in each node ({node: step})
freeNodes = list(range(1, self.numberOfProcs+1)) # available nodes to send jobs
logger.info("Execution threads: %s" % freeNodes)
logger.info("Running steps using %s threads. 1 thread is used for this main proccess." % self.numberOfProcs)
while True:
# See which of the runningSteps are not really running anymore.
# Update them and freeNodes, and call final callback for step.
with sharedLock:
nodesFinished = [node for node, step in runningSteps.items()
if not step.isRunning()]
doContinue = True
for node in nodesFinished:
step = runningSteps.pop(node) # remove entry from runningSteps
freeNodes.append(node) # the node is available now
self.freeGpusSlot(step.getObjId())
# Notify steps termination and check if we should continue
doContinue = stepFinishedCallback(step)
if not doContinue:
break
if not doContinue:
break
anyLaunched = False
# If there are available nodes, send next runnable step.
with sharedLock:
if freeNodes:
runnableSteps = self._getRunnable(steps, len(freeNodes))
for step in runnableSteps:
# We found a step to work in, so let's start a new
# thread to do the job and book it.
anyLaunched = True
step.setRunning()
stepStartedCallback(step)
node = freeNodes.pop(0) # take an available node
runningSteps[node] = step
logger.debug("Running step %s on node %s" % (step, node))
t = StepThread(step, sharedLock)
# won't keep process up if main thread ends
t.daemon = True
t.start()
anyPending = self._arePending(steps)
if not anyLaunched:
if anyPending: # nothing running
time.sleep(3)
else:
break # yeah, we are done, either failed or finished :)
now = datetime.datetime.now()
if now - lastCheck > delta:
stepsCheckCallback()
lastCheck = now
stepsCheckCallback()
# Wait for all threads now.
for t in threading.enumerate():
if t is not threading.current_thread():
t.join()
[docs]class QueueStepExecutor(ThreadStepExecutor):
def __init__(self, hostConfig, submitDict, nThreads, **kwargs):
ThreadStepExecutor.__init__(self, hostConfig, nThreads, **kwargs)
self.submitDict = submitDict
# Command counter per thread
self.threadCommands = {}
if nThreads > 1:
self.runJobs = ThreadStepExecutor.runSteps
else:
self.runJobs = StepExecutor.runSteps
self.renameGpuIds()
[docs] def renameGpuIds(self):
""" Reorganize the gpus ids starting from 0 since the queue engine is the one assigning them.
https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars """
for threadId, gpuList in self.gpuDict.items():
for i in range(len(gpuList)):
self.gpuDict[threadId][i] = i
logger.debug("Updated gpus ids rebase starting from 0: %s per thread" %self.gpuDict)
[docs] def getThreadJobId(self, stepId):
""" Returns the job id extension assigned to each thread/step """
if not stepId in self.threadCommands:
self.threadCommands[stepId] = 0
self.threadCommands[stepId] += 1
return self.threadCommands[stepId]
[docs] def runJob(self, log, programName, params, numberOfMpi=1, numberOfThreads=1, env=None, cwd=None, executable=None):
threadId = threading.current_thread().thId
submitDict = dict(self.hostConfig.getQueuesDefault())
submitDict.update(self.submitDict)
submitDict['JOB_COMMAND'] = process.buildRunCommand(programName, params, numberOfMpi,
self.hostConfig, env,
gpuList=self.getGpuList())
threadJobId = self.getThreadJobId(threadId)
subthreadId = '-%s-%s' % (threadId, threadJobId)
submitDict['JOB_NAME'] = submitDict['JOB_NAME'] + subthreadId
submitDict['JOB_SCRIPT'] = os.path.abspath(removeExt(submitDict['JOB_SCRIPT']) + subthreadId + ".job")
submitDict['JOB_LOGS'] = os.path.join(getParentFolder(submitDict['JOB_SCRIPT']), submitDict['JOB_NAME'])
jobid = _submit(self.hostConfig, submitDict, cwd, env)
self.protocol.appendJobId(jobid) # append active jobs
self.protocol._store(self.protocol._jobId)
if (jobid is None) or (jobid == UNKNOWN_JOBID):
logger.info("jobId is none therefore we set it to fail")
raise Exception("Failed to submit to queue.")
status = cts.STATUS_RUNNING
wait = 3
# Check status while job running
# REVIEW this to minimize the overhead in time put by this delay check
while _checkJobStatus(self.hostConfig, jobid) == cts.STATUS_RUNNING:
time.sleep(wait)
if wait < 300:
wait += 3
self.protocol.removeJobId(jobid) # After completion, remove inactive jobs.
self.protocol._store(self.protocol._jobId)
return status