""" This module has utilities to benchmark paraview. First, when run standalone, this will do a simple rendering benchmark test. The test renders a sphere with various rendering settings and reports the rendering rate achieved in triangles/sec. run() is the entrypoint for that usage. Second, you can set up arbitrary pipelines and this module helps you obtain, interpret and report the information recorded by ParaView's logs. Do that like so: 1. optionally, call maximize logs first 2. setup and run your visualization pipeline (via GUI or script as you prefer) 3. either call print_logs() to print out the logs in raw format call parse_logs() to let the script identify and report on per frame and per filter execution times :: WARNING: This was meant for server side rendering, but it could work reasonably well when geometry is delivered to the client and rendered there if the script were changed to recognize MPIMoveData as end of frame and did something sensible on the server which has no other end of frame knowledge TODO: builtin mode shouldn't show server info, it is redundant TODO: this doesn't handle split render/data server mode TODO: the end of frame markers are heuristic, likely buggy, and have not been tried since before 3.9's view restructuring """ import time import sys from paraview.simple import * try: import numpy numpy_loaded = True except ImportError: numpy_loaded = False import re import paraview import copy import pickle # a regular expression to parse filter execution time match_filter = re.compile(" *Execute (\w+) id: +(\d+), +(\d*.*\d+) +seconds") match_vfilter = re.compile(" *Execute (\w+) +, +(\d*.*\d+) +seconds") # a regular expression to parse overall rendering time match_still_render = re.compile(" *(Still) Render, +(\d*.*\d+) +seconds") match_interactive_render = \ re.compile(" *(Interactive) Render, +(\d*.*\d+) +seconds") match_render = re.compile(" *(\w+|\w+ Dev) Render, +(\d*.*\d+) +seconds") match_icetrender = re.compile("(IceT Dev) Render, +(\d*.*\d+) +seconds") # more for parallel composite and delivery time match_composite = re.compile(" *Compositing, +(\d*.*\d+) +seconds") match_send = re.compile(" *Sending, +(\d*.*\d+) +seconds") match_receive = re.compile(" *Receiving, +(\d*.*\d+) +seconds") match_comp_xmit = \ re.compile(" *TreeComp (Send|Receive) (\d+) " + \ "(to|from) (\d+) uchar (\d+), +(\d*.*\d+) +seconds") match_comp_comp = re.compile(" *TreeComp composite, *(\d*.*\d+) +seconds") showparse = False #icet composite message comes after the render messages, #where for bswap and manta it comes before so we have to treat icet differently icetquirk = False start_frame = 0 default_log_threshold = dict() class OneLog : def __init__(self): self.runmode = 'batch' self.servertype = 'unified' self.component = 0x10 self.rank = 0 self.lines = [] def componentString(self): ret = "" if self.component & 0x10: ret = ret + " CLIENT " if self.component & 0x4: ret = ret + " RENDER " if self.component & 0x1: ret = ret + " DATA " return ret def print_log(self, showlines=False): print "#RunMode:", self.runmode, print "ServerType:", self.servertype, print "Component:", self.componentString(), print "processor#:", self.rank if showlines: for i in self.lines: print i def toString(self, showlines=False): result = "#RunMode: " + self.runmode + " ServerType: " + self.servertype + " Component: " + self.componentString() + " processor#: " + str(self.rank) + "\n" if showlines: for i in self.lines: result += i + "\n" return result logs = [] def maximize_logs () : """ Convenience method to ask paraview to produce logs with lots of space and highest resolution. """ pm = paraview.servermanager.vtkProcessModule.GetProcessModule() if pm == None: return ss = paraview.servermanager.vtkSMSession for ptype in [ss.CLIENT_AND_SERVERS, ss.CLIENT, ss.SERVERS, ss.RENDER_SERVER, ss.DATA_SERVER]: default_log_threshold[str(ptype)] = 0.0 pxm = paraview.servermanager.ProxyManager() tl = pxm.NewProxy("misc", "TimerLog") prop = tl.GetProperty("MaxEntries") prop.SetElements1(1000000) tl.UpdateVTKObjects() def get_memuse() : session = servermanager.ProxyManager().GetSessionProxyManager().GetSession() retval = [] infos = servermanager.vtkPVMemoryUseInformation() session.GatherInformation(session.CLIENT, infos, 0) procUse = str(infos.GetProcMemoryUse(0)) hostUse = str(infos.GetHostMemoryUse(0)) retval.append("CLIENT " + procUse + " / " + hostUse) infos = servermanager.vtkPVMemoryUseInformation() session.GatherInformation(session.DATA_SERVER, infos, 0) for i in range(0,infos.GetSize()): rank = str(infos.GetRank(i)) procUse = str(infos.GetProcMemoryUse(i)) hostUse = str(infos.GetHostMemoryUse(i)) retval.append("DS[" + rank + "] " + procUse + " / " + hostUse) return retval def dump_logs( filename ) : """ This saves off the logs we've gathered. Ot allows you to run a benchmark somewhere, save off all of the details in raw format, then load them somewhere else. You can then do a detailed analysis and you always have the raw data to go back to. """ global logs f = open(filename, "w") pickle.dump(logs, f) f.close() def import_logs( filename ) : """ This is for bringing in a saved log files and parse it after the fact. TODO: add an option to load in raw parview logs in text format """ global logs logs = [] f = open(filename, "r") logs = pickle.load(f) f.close() def get_logs() : """ This is for bringing in logs at run time to parse while running. """ global logs logs = [] pm = paraview.servermanager.vtkProcessModule.GetProcessModule() if pm == None: return connectionId = paraview.servermanager.ActiveConnection.ID session = paraview.servermanager.ActiveConnection.Session is_symmetric_mode = False if pm.GetProcessTypeAsInt() == pm.PROCESS_BATCH: runmode = 'batch' is_symmetric_mode = pm.GetSymmetricMPIMode() else: runmode = 'interactive' if session.GetRenderClientMode() == session.RENDERING_UNIFIED: servertype = 'unified' else: servertype = 'split' if runmode == 'batch': # collect information from all processes in one go. components = [session.CLIENT_AND_SERVERS] else: if servertype == 'unified': # collect information separately for client and servers. components = [session.CLIENT, session.SERVERS] else: # collect information separately for all process types. components = [session.CLIENT, session.RENDER_SERVER, session.DATA_SERVER] for component in components: timerInfo = paraview.servermanager.vtkPVTimerInformation() if len(default_log_threshold) != 0: timerInfo.SetLogThreshold(default_log_threshold[str(component)]) session.GatherInformation(component, timerInfo, 0) for i in range(timerInfo.GetNumberOfLogs()): alog = OneLog() alog.runmode = runmode alog.servertype = servertype alog.component = component alog.rank = i if is_symmetric_mode: # in Symmetric mode, GatherInformation() only collects # information from the current node. so the # vtkPVTimerInformation will only have info for local process. alog.rank = pm.GetPartitionId() for line in timerInfo.GetLog(i).split('\n'): alog.lines.append(line) logs.append(alog) def print_logs() : """ Print logs on the root node by gathering logs accross all the nodes regardless if the process was started in symmetric mode or not. """ global logs if len(logs) == 0: get_logs() # Handle symetric mode specificaly if need be pm = paraview.servermanager.vtkProcessModule.GetProcessModule() is_symmetric_mode = False if pm != None: is_symmetric_mode = pm.GetSymmetricMPIMode() if is_symmetric_mode: # Need to provide extra synchronization ctrl = pm.GetGlobalController() proc = pm.GetPartitionId() nbProc = pm.GetNumberOfLocalPartitions() if proc == 0: # Start with my logs for i in logs: i.print_log(True) # Then Print the log of every other rank for otherProc in range(1, nbProc): # Max buffer size 999999 logSize = " " * 6 ctrl.Receive(logSize, len(logSize), otherProc, 987455) logSize = int(logSize) logTxt = " " * logSize ctrl.Receive(logTxt, logSize, otherProc, 987456) print logTxt else: # Extract logs text logTxt = "" for i in logs: logTxt += i.toString(True) logSize = str(len(logTxt)) # Push local logs to process 0 ctrl.Send(logSize, len(logSize), 0, 987455) ctrl.Send(logTxt, len(logTxt), 0, 987456) else: # Regular local print for i in logs: i.print_log(True) def __process_frame() : global filters global current_frames_records global frames global start_frame max = len(current_frames_records) #determine ancestry of each record from order and indent #subtract only immediate children from each record #TODO: Make this an option for x in xrange(max): indent = current_frames_records[x]['indent'] minindent = 10000 for y in xrange(x+1,max): indent2 = current_frames_records[y]['indent'] if indent2<=indent: #found a record which is not a descendant break if indent2 < minindent: minindent = indent2 for y in xrange(x+1,max): indent2 = current_frames_records[y]['indent'] if indent2 == minindent: current_frames_records[x]['local_duration'] = \ current_frames_records[x]['local_duration'] -\ current_frames_records[y]['duration'] for x in xrange(max): #keep global statics per filter record = current_frames_records[x] id = record['id'] if id in filters: srecord = filters[id] srecord['duration'] = srecord['duration'] + record['duration'] srecord['local_duration'] = srecord['local_duration'] +\ record['local_duration'] srecord['count'] = srecord['count'] + 1 filters[id] = srecord else: filters[id] = copy.deepcopy(record) #save off this frame and begin the next frames.append(current_frames_records) current_frames_records = [] def __parse_line (line) : """ Examine one line from the logs. If it is a report about a filter's execution time, parse the relevant information out of the line and collect those statistics. We record each filter's average execution time as well as the each filters contribution to the each rendered frame. """ global filters global current_frames_records global cnt global show_input global icetquirk found = False #find indent cnt = 0 for c in range(len(line)): if line[c] == " ": cnt = cnt + 1 else: break #determine if this log comes from icet so we can #do special case treatement for frame markings icetline = False match = match_icetrender.match(line) if match != None: icetquirk = True icetline = True match = match_filter.match(line) if match != None: found = True if showparse: print "FILT:", cnt, line name = match.group(1) id = match.group(2) duration = match.group(3) match = match_vfilter.match(line) if match != None: found = True if showparse: print "LFLT:", cnt, line name = match.group(1) id = name duration = match.group(2) match = match_comp_comp.match(line) if match != None: found = True if showparse: print "TCMP:", cnt, line name = "tree comp" id = name duration = match.group(1) match = match_comp_xmit.match(line) if match != None: found = True if showparse: print "TXMT:", cnt, line name = match.group(1) id = name duration = match.group(6) match = match_composite.match(line) if match != None: found = True if showparse: print "COMP:", cnt, line name = 'composite' id = 'comp' duration = match.group(1) match = match_send.match(line) if match != None: found = True if showparse: print "SEND:", cnt, line name = 'send' id = 'send' duration = match.group(1) match = match_receive.match(line) if match != None: found = True if showparse: print "RECV:", cnt, line name = 'receive' id = 'recv' duration = match.group(1) match = match_still_render.match(line) if match != None: found = True if showparse: print "STILL:", cnt, line name = match.group(1) id = 'still' duration = match.group(2) if match == None: match = match_interactive_render.match(line) if match != None: found = True if showparse: print "INTER:", cnt, line name = match.group(1) id = 'inter' duration = match.group(2) if match == None: match = match_render.match(line) if match != None: found = True if showparse: print "REND:", cnt, line name = match.group(1) id = 'render' duration = match.group(2) if found == False: # we didn't find anything we recognized in this line, ignore it if showparse: print "????:", cnt, line return record = dict() record['id'] = id record['name'] = name record['duration'] = float(duration) record['local_duration'] = float(duration) record['count'] = 1 record['indent'] = cnt #watch for the beginning of the next frame/end of previous frame if cnt == 0: if (id == 'still') or \ (id == 'inter') or \ (icetquirk == False and id == 'comp') or \ (icetquirk == True and icetline == True) : if showparse: print "SOF" #start of frame #decipher parent child information from records in the frame #and save off newly gathered per filter and per frame statistics __process_frame() #keep a record of this execution as part for the current frame current_frames_records.append(record) return def parse_logs(show_parse = False, tabular = False) : """ Parse the collected paraview log information. This prints out per frame, and aggregated per filter statistics. If show_parse is true, debugging information is shown about the parsing process that allows you to verify that the derived stats are correct. This includes each and echo of each log line collected, prepended by the token type and indent scanned in, or ???? if the line is unrecognized and ignored. Frame boundaries are denoted by SOF, indicating the preceeding line was determined to be the start of the next frame. """ global filters global current_frames_records global frames global cnt global showparse global start_frame showparse = show_parse if len(logs) == 0: get_logs() for i in logs: # per filter records filters = dict() filters.clear() # per frame records frames = [] # components of current frame current_frames_records = [] cnt = 0 runmode = i.runmode servertype = i.servertype component = i.component rank = i.rank i.print_log(False) for line in i.lines: __parse_line(line) #collect stats for the current frame in process but not officially ended __process_frame() #print out the gathered per frame information if tabular: frecs = dict() line = "#framenum, " for x in filters: line += filters[x]['name'] + ":" + filters[x]['id'] + ", " #print line for cnt in xrange(start_frame, len(frames)): line = "" line += str(cnt) + ", " printed = dict() for x in filters: id = filters[x]['id'] name = filters[x]['name'] found = False for record in frames[cnt]: if 'id' in record: if record['id'] == id and \ record['name'] == name and \ not id in printed: found = True printed[id] = 1 line += str(record['local_duration']) + ", " if not id in frecs: frecs[id] = [] frecs[id].append(record['local_duration']) if not found: line += "0, " #print line #print for x in frecs.keys(): v = frecs[x] print "# ", x, len(v), if numpy_loaded: print numpy.min(v), numpy.mean(v), numpy.max(v), print numpy.std(v) else: print "#FRAME TIMINGS" print "#filter id, filter type, inclusive duration, local duration" for cnt in xrange(start_frame, len(frames)): print "#Frame ", cnt for record in frames[cnt]: if 'id' in record: print record['id'], ",", print record['name'], ",", print record['duration'], ",", print record['local_duration'] #print #print if not tabular: #print out the gathered per filter information print "#FILTER TIMINGS" print "#filter id, filter type, count, "+\ "sum inclusive duration, sum local duration" for x in filters: record = filters[x] print record['id'], ",", print record['name'], ",", print record['count'], ",", print record['duration'], ",", print record['local_duration'] print def __render(ss, v, title, nframes): print '============================================================' print title res = [] res.append(title) for phires in (500, 1000): ss.PhiResolution = phires c = v.GetActiveCamera() v.CameraPosition = [-3, 0, 0] v.CameraFocalPoint = [0, 0, 0] v.CameraViewUp = [0, 0, 1] Render() c1 = time.time() for i in range(nframes): c.Elevation(0.5) Render() tpr = (time.time() - c1)/nframes ncells = ss.GetDataInformation().GetNumberOfCells() print tpr, " secs/frame" print ncells, " polys" print ncells/tpr, " polys/sec" res.append((ncells, ncells/tpr)) return res def run(filename=None, nframes=60): """ Runs the benchmark. If a filename is specified, it will write the results to that file as csv. The number of frames controls how many times a particular configuration is rendered. Higher numbers lead to more accurate averages. """ # Turn off progress printing paraview.servermanager.SetProgressPrintingEnabled(0) # Create a sphere source to use in the benchmarks ss = Sphere(ThetaResolution=1000, PhiResolution=500) rep = Show() v = Render() results = [] # Start with these defaults #v.RemoteRenderThreshold = 0 obj = servermanager.misc.GlobalMapperProperties() obj.GlobalImmediateModeRendering = 0 # Test different configurations title = 'display lists, no triangle strips, solid color' obj.GlobalImmediateModeRendering = 0 results.append(__render(ss, v, title, nframes)) title = 'no display lists, no triangle strips, solid color' obj.GlobalImmediateModeRendering = 1 results.append(__render(ss, v, title, nframes)) # Color by normals lt = servermanager.rendering.PVLookupTable() rep.LookupTable = lt rep.ColorArrayName = "Normals" lt.RGBPoints = [-1, 0, 0, 1, 0.0288, 1, 0, 0] lt.ColorSpace = 'HSV' lt.VectorComponent = 0 title = 'display lists, no triangle strips, color by array' obj.GlobalImmediateModeRendering = 0 results.append(__render(ss, v, title, nframes)) title = 'no display lists, no triangle strips, color by array' obj.GlobalImmediateModeRendering = 1 results.append(__render(ss, v, title, nframes)) if filename: f = open(filename, "w") else: f = sys.stdout print >>f, 'configuration, %d, %d' % (results[0][1][0], results[0][2][0]) for i in results: print >>f, '"%s", %g, %g' % (i[0], i[1][1], i[2][1]) def test_module(): """Simply exercises a few components of the module.""" maximize_logs() paraview.servermanager.SetProgressPrintingEnabled(0) ss = Sphere(ThetaResolution=1000, PhiResolution=500) rep = Show() v = Render() print_logs() if __name__ == "__main__": if "--test" in sys.argv: test_module() else: run()