Logo ROOT   6.14/05
Reference Guide
TProofCondor.cxx
Go to the documentation of this file.
1 // @(#)root/proof:$Id$
2 // Author: Fons Rademakers 13/02/97
3 
4 /*************************************************************************
5  * Copyright (C) 1995-2000, Rene Brun and Fons Rademakers. *
6  * All rights reserved. *
7  * *
8  * For the licensing terms see $ROOTSYS/LICENSE. *
9  * For the list of contributors see $ROOTSYS/README/CREDITS. *
10  *************************************************************************/
11 
12 //////////////////////////////////////////////////////////////////////////
13 // //
14 // TProof //
15 // //
16 // This class controls a Parallel ROOT Facility, PROOF, cluster. //
17 // It fires the slave servers, it keeps track of how many slaves are //
18 // running, it keeps track of the slaves running status, it broadcasts //
19 // messages to all slaves, it collects results, etc. //
20 // //
21 //////////////////////////////////////////////////////////////////////////
22 
23 #include "TProofCondor.h"
24 
25 #include "TCondor.h"
26 #include "TList.h"
27 #include "TMap.h"
28 #include "TMessage.h"
29 #include "TMonitor.h"
30 #include "TProofNodeInfo.h"
31 #include "TProofResourcesStatic.h"
32 #include "TProofServ.h"
33 #include "TSlave.h"
34 #include "TSocket.h"
35 #include "TString.h"
36 #include "TTimer.h"
37 
39 
40 ////////////////////////////////////////////////////////////////////////////////
41 /// Start proof using condor
42 
43 TProofCondor::TProofCondor(const char *masterurl, const char *conffile,
44  const char *confdir, Int_t loglevel,
45  const char *, TProofMgr *mgr)
46  : fCondor(0), fTimer(0)
47 {
48  // Default initializations
49  InitMembers();
50 
51  // This may be needed during init
52  fManager = mgr;
53 
54  fUrl = TUrl(masterurl);
55 
56  if (!conffile || !conffile[0]) {
57  conffile = kPROOF_ConfFile;
58  } else if (!strncasecmp(conffile, "condor:", 7)) {
59  conffile+=7;
60  }
61 
62  if (!confdir || !confdir[0]) {
63  confdir = kPROOF_ConfDir;
64  }
65 
66  Init(masterurl, conffile, confdir, loglevel);
67 }
68 
69 ////////////////////////////////////////////////////////////////////////////////
70 /// Clean up Condor PROOF environment.
71 
73 {
76 }
77 
78 ////////////////////////////////////////////////////////////////////////////////
79 /// Setup Condor workers using dynamic information
80 
82 {
83  fCondor = new TCondor;
84  TString jobad = GetJobAd();
85 
87  if (fImage.Length() == 0) {
88  Error("StartSlaves", "Empty Condor image found for system %s",
89  gSystem->HostName());
90  return kFALSE;
91  }
92 
93  TList claims;
94  if (fConfFile.IsNull()) {
95  // startup all slaves if no config file given
96  TList *condorclaims = fCondor->Claim(9999, jobad);
97  TIter nextclaim(condorclaims);
98  while (TObject *o = nextclaim()) claims.Add(o);
99  } else {
100  // parse config file
102  fConfFile = resources->GetFileName(); // Update the global file name (with path)
103  PDB(kGlobal,1) Info("StartSlaves", "using PROOF config file: %s", fConfFile.Data());
104 
105  // Get all workers
106  TList *workerList = resources->GetWorkers();
107  if (workerList->GetSize() == 0) {
108  Error("StartSlaves", "Found no condorworkers in %s", fConfFile.Data());
109  return kFALSE;
110  }
111 
112  // check for valid slave lines and claim condor nodes
113  Int_t ord = 0;
114 
115  // Loop over all workers and start them
116  TListIter next(workerList);
117  TObject *to;
118  TProofNodeInfo *worker;
119  int nSlavesDone = 0;
120  while ((to = next())) {
121  // Get the next worker from the list
122  worker = (TProofNodeInfo *)to;
123 
124  // Read back worker node info
125  const Char_t *image = worker->GetImage().Data();
126  const Char_t *workdir = worker->GetWorkDir().Data();
127  Int_t perfidx = worker->GetPerfIndex();
128 
129  gSystem->Sleep(10 /* ms */);
130  TCondorSlave* csl = fCondor->Claim(worker->GetNodeName().Data(), jobad);
131  if (csl) {
132  csl->fPerfIdx = perfidx;
133  csl->fImage = image;
134  csl->fWorkDir = gSystem->ExpandPathName(workdir);
135  TString fullord = TString(gProofServ->GetOrdinal()) + "." + ((Long_t) ord);
136  csl->fOrdinal = fullord.Data();
137  claims.Add(csl);
138  ord++;
139  }
140 
141  // Notify claim creation
142  nSlavesDone++;
144  m << TString("Creating COD Claim") << workerList->GetSize()
145  << nSlavesDone << (csl != 0);
146  gProofServ->GetSocket()->Send(m);
147 
148  } // end while (worker loop)
149 
150  // Cleanup
151  delete resources;
152  resources = 0;
153  } // end else (parse config file)
154 
155  Long_t delay = 500; // timer delay 0.5s
156  Int_t ntries = 20; // allow 20 tries (must be > 1 for algorithm to work)
157  Int_t trial = 1;
158  Int_t idx = 0;
159 
160  int nClaims = claims.GetSize();
161  int nClaimsDone = 0;
162  while (claims.GetSize() > 0) {
163  TCondorSlave* c = 0;
164 
165  // Get Condor Slave
166  if (trial == 1) {
167  c = dynamic_cast<TCondorSlave*>(claims.At(idx));
168  } else {
169  TPair *p = dynamic_cast<TPair*>(claims.At(idx));
170  if (p) {
171  TTimer *t = dynamic_cast<TTimer*>(p->Value());
172  if (t) {
173  // wait remaining time
174  Long64_t wait = t->GetAbsTime()-gSystem->Now();
175  if (wait > 0) gSystem->Sleep((UInt_t)wait);
176  c = dynamic_cast<TCondorSlave*>(p->Key());
177  }
178  }
179  }
180 
181  // create slave
182  TSlave *slave = 0;
183  if (c) slave = CreateSlave(Form("%s:%d", c->fHostname.Data(), c->fPort), c->fOrdinal,
184  c->fPerfIdx, c->fImage, c->fWorkDir);
185 
186  // add slave to appropriate list
187  if (trial < ntries) {
188  if (slave && slave->IsValid()) {
189  fSlaves->Add(slave);
190  if (trial == 1) {
191  claims.Remove(c);
192  } else {
193  TPair *p = dynamic_cast<TPair*>(claims.Remove(c));
194  if (p) {
195  TTimer *xt = dynamic_cast<TTimer*>(p->Value());
196  if (xt) delete xt;
197  delete p;
198  }
199  }
200  nClaimsDone++;
202  m << TString("Opening connections to workers") << nClaims
203  << nClaimsDone << kTRUE;
204  gProofServ->GetSocket()->Send(m);
205  } else if (slave) {
206  if (trial == 1) {
207  TTimer* timer = new TTimer(delay);
208  TPair *p = new TPair(c, timer);
209  claims.RemoveAt(idx);
210  claims.AddAt(p, idx);
211  } else {
212  TPair *p = dynamic_cast<TPair*>(claims.At(idx));
213  if (p && p->Value()) {
214  TTimer *xt = dynamic_cast<TTimer*>(p->Value());
215  if (xt) xt->Reset();
216  }
217  }
218  delete slave;
219  idx++;
220  } else {
221  Warning("StartSlaves", "could not create TSlave object!");
222  }
223  } else {
224  if (slave) {
225  fSlaves->Add(slave);
226  TPair *p = dynamic_cast<TPair*>(claims.Remove(c));
227  if (p && p->Value()) {
228  TTimer *xt = dynamic_cast<TTimer*>(p->Value());
229  delete xt;
230  }
231  if (p) delete p;
232 
233  nClaimsDone++;
235  m << TString("Opening connections to workers") << nClaims
236  << nClaimsDone << slave->IsValid();
237  gProofServ->GetSocket()->Send(m);
238  } else {
239  Warning("StartSlaves", "could not create TSlave object!");
240  }
241  }
242 
243  if (idx>=claims.GetSize()) {
244  trial++;
245  idx = 0;
246  }
247  }
248 
249  // Here we finalize the server startup: in this way the bulk
250  // of remote operations are almost parallelized
251  TIter nxsl(fSlaves);
252  TSlave *sl = 0;
253  int nSlavesDone = 0, nSlavesTotal = fSlaves->GetSize();
254  while ((sl = (TSlave *) nxsl())) {
255 
256  // Finalize setup of the server
257  if (sl->IsValid()) {
258  sl->SetupServ(TSlave::kSlave, 0);
259  }
260 
261  if (sl->IsValid()) {
262  fAllMonitor->Add(sl->GetSocket());
263  } else {
264  fBadSlaves->Add(sl);
265  }
266 
267  // Notify end of startup operations
268  nSlavesDone++;
270  Bool_t wrkvalid = sl->IsValid() ? kTRUE : kFALSE;
271  m << TString("Setting up worker servers") << nSlavesTotal
272  << nSlavesDone << wrkvalid;
273  gProofServ->GetSocket()->Send(m);
274  }
275 
276  return kTRUE;
277 }
278 
279 ////////////////////////////////////////////////////////////////////////////////
280 /// Suspend or resume PROOF via Condor.
281 
283 {
284  if (fTimer == 0) {
285  fTimer = new TTimer();
286  }
287  if (active) {
288  PDB(kCondor,1) Info("SetActive","-- Condor Resume --");
289  fTimer->Stop();
291  fCondor->Resume();
292  } else {
293 #if 1
294  return; // don't suspend for the moment
295 #else
296  Int_t delay = 60000; // milli seconds
297  PDB(kCondor,1) Info("SetActive","-- Delayed Condor Suspend (%d msec / to %lld) --",
298  delay, delay + Long64_t(gSystem->Now()));
299  fTimer->Connect("Timeout()", "TCondor", fCondor, "Suspend()");
300  fTimer->Start(10000, kTRUE); // single shot
301 #endif
302  }
303 }
304 
305 ////////////////////////////////////////////////////////////////////////////////
306 /// Get job Ad
307 
309 {
310  TString ad;
311 
312  ad = "JobUniverse = 5\n"; // vanilla
313  ad += Form("Cmd = \"%s/bin/proofd\"\n", GetConfDir());
314  ad += Form("Iwd = \"%s\"\n", gSystem->TempDirectory());
315  ad += "In = \"/dev/null\"\n";
316  ad += Form("Out = \"%s/proofd.out.$(Port)\"\n", gSystem->TempDirectory());
317  ad += Form("Err = \"%s/proofd.err.$(Port)\"\n", gSystem->TempDirectory());
318  ad += Form("Args = \"-f -p $(Port) -d %d %s\"\n", GetLogLevel(), GetConfDir());
319 
320  return ad;
321 }
virtual void AddAt(TObject *obj, Int_t idx)
Insert object at position idx in the list.
Definition: TList.cxx:303
virtual void Info(const char *method, const char *msgfmt,...) const
Issue info message.
Definition: TObject.cxx:854
long long Long64_t
Definition: RtypesCore.h:69
auto * m
Definition: textangle.C:8
TString fConfDir
Definition: TProof.h:569
void Reset()
Reset the timer.
Definition: TTimer.cxx:157
EState GetState() const
Definition: TCondor.h:79
virtual Int_t Send(const TMessage &mess)
Send a TMessage object.
Definition: TSocket.cxx:527
This class represents a WWW compatible URL.
Definition: TUrl.h:35
TString fWorkDir
Definition: TCondor.h:42
TString GetImage(const char *host) const
Get image of the worker.
Definition: TCondor.cxx:439
The PROOF manager interacts with the PROOF server coordinator to create or destroy a PROOF session...
Definition: TProofMgr.h:43
TSocket * GetSocket() const
Definition: TProofServ.h:257
TString fImage
Definition: TProof.h:570
Int_t GetPerfIndex() const
virtual void Add(TSocket *sock, Int_t interest=kRead)
Add socket to the monitor&#39;s active list.
Definition: TMonitor.cxx:168
Basic string class.
Definition: TString.h:131
int Int_t
Definition: RtypesCore.h:41
bool Bool_t
Definition: RtypesCore.h:59
const TString & GetImage() const
Iterator of linked list.
Definition: TList.h:197
virtual TObject * RemoveAt(Int_t idx)
const char * GetConfDir() const
Definition: TProof.h:904
TTime GetAbsTime() const
Definition: TTimer.h:78
TList * fBadSlaves
Definition: TProof.h:574
TList * GetWorkers()
Get the list of worker nodes.
virtual Bool_t IsValid() const
Definition: TSlave.h:150
#define PDB(mask, level)
Definition: TProofDebug.h:56
virtual void Sleep(UInt_t milliSec)
Sleep milliSec milli seconds.
Definition: TSystem.cxx:445
TSlave * CreateSlave(const char *url, const char *ord, Int_t perf, const char *image, const char *workdir)
Create a new TSlave of type TSlave::kSlave.
Definition: TProof.cxx:1831
TSocket * GetSocket() const
Definition: TSlave.h:134
Bool_t StartSlaves(Bool_t)
Setup Condor workers using dynamic information.
Int_t GetLogLevel() const
Definition: TProof.h:916
virtual void Start(Long_t milliSec=-1, Bool_t singleShot=kFALSE)
Starts the timer with a milliSec timeout.
Definition: TTimer.cxx:211
TList * Claim(Int_t n, const char *cmd)
Claim n virtual machines This function figures out the image and performance index before returning t...
Definition: TCondor.cxx:235
Int_t Init(const char *masterurl, const char *conffile, const char *confdir, Int_t loglevel, const char *alias=0)
Start the PROOF environment.
Definition: TProof.cxx:745
TString fConfFile
Definition: TProof.h:568
TList * fSlaves
Definition: TProof.h:572
virtual const char * TempDirectory() const
Return a user configured or systemwide directory to create temporary files in.
Definition: TSystem.cxx:1458
virtual void SetActive()
Definition: TProofCondor.h:50
TObject * Value() const
Definition: TMap.h:121
A doubly linked list.
Definition: TList.h:44
const char *const kPROOF_ConfFile
Definition: TProof.h:122
TMonitor * fAllMonitor
Definition: TProof.h:575
void InitMembers()
Default initializations.
Definition: TProof.cxx:524
Bool_t Connect(const char *signal, const char *receiver_class, void *receiver, const char *slot)
Non-static method is used to connect from the signal of this object to the receiver slot...
Definition: TQObject.cxx:867
TString fHostname
Definition: TCondor.h:36
TProofMgr * fManager
Definition: TProof.h:587
virtual TTime Now()
Get current time in milliseconds since 0:00 Jan 1 1995.
Definition: TSystem.cxx:471
The purpose of this class is to provide a complete node description for masters, submasters and worke...
const TString & GetNodeName() const
R__EXTERN TSystem * gSystem
Definition: TSystem.h:540
TObject * Key() const
Definition: TMap.h:120
TProofCondor(const char *masterurl, const char *conffile=kPROOF_ConfFile, const char *confdir=kPROOF_ConfDir, Int_t loglevel=0, const char *alias=0, TProofMgr *mgr=0)
Start proof using condor.
virtual TObject * Remove(TObject *obj)
Remove object from the list.
Definition: TList.cxx:818
const char *const kPROOF_ConfDir
Definition: TProof.h:123
unsigned int UInt_t
Definition: RtypesCore.h:42
virtual void Error(const char *method, const char *msgfmt,...) const
Issue error message.
Definition: TObject.cxx:880
char * Form(const char *fmt,...)
Ssiz_t Length() const
Definition: TString.h:405
virtual ~TProofCondor()
Clean up Condor PROOF environment.
TTimer * fTimer
Definition: TProofCondor.h:39
Handles synchronous and a-synchronous timer events.
Definition: TTimer.h:51
virtual TObject * At(Int_t idx) const
Returns the object at position idx. Returns 0 if idx is out of range.
Definition: TList.cxx:354
const Bool_t kFALSE
Definition: RtypesCore.h:88
TString GetJobAd()
Get job Ad.
TString fOrdinal
Definition: TCondor.h:41
#define SafeDelete(p)
Definition: RConfig.h:529
long Long_t
Definition: RtypesCore.h:50
Class used by TMap to store (key,value) pairs.
Definition: TMap.h:102
#define ClassImp(name)
Definition: Rtypes.h:359
virtual const char * HostName()
Return the system&#39;s host name.
Definition: TSystem.cxx:311
TCondor * fCondor
Definition: TProofCondor.h:38
Bool_t IsNull() const
Definition: TString.h:402
Mother of all ROOT objects.
Definition: TObject.h:37
The purpose of this class is to provide a standard interface to static config files.
char Char_t
Definition: RtypesCore.h:29
R__EXTERN TProofServ * gProofServ
Definition: TProofServ.h:347
const TString & GetWorkDir() const
virtual void Add(TObject *obj)
Definition: TList.h:87
Int_t fPort
Definition: TCondor.h:37
friend class TCondor
Definition: TProofCondor.h:35
#define c(i)
Definition: RSha256.hxx:101
virtual void Stop()
Definition: TTimer.h:93
virtual Bool_t ExpandPathName(TString &path)
Expand a pathname getting rid of special shell characters like ~.
Definition: TSystem.cxx:1254
virtual Int_t GetSize() const
Return the capacity of the collection, i.e.
Definition: TCollection.h:182
Class describing a PROOF worker server.
Definition: TSlave.h:46
TUrl fUrl
Definition: TProof.h:567
const Bool_t kTRUE
Definition: RtypesCore.h:87
Bool_t Resume()
Resume worker.
Definition: TCondor.cxx:342
const char * GetOrdinal() const
Definition: TProofServ.h:253
virtual Int_t SetupServ(Int_t stype, const char *conffile)
Init a PROOF slave object.
Definition: TSlave.cxx:179
TString fImage
Definition: TCondor.h:39
virtual void Warning(const char *method, const char *msgfmt,...) const
Issue warning message.
Definition: TObject.cxx:866
const char * Data() const
Definition: TString.h:364
Int_t fPerfIdx
Definition: TCondor.h:38