ndb - Add possibility to limit disk write speed in backup(lcp)

      3 new paramters:
      DiskSyncSize - Outstanding disk writes before sync (default 4M)
      DiskCheckpointSpeed - Write speed of LCP in bytes/sec (default 10M)
      DiskCheckpointSpeedInRestart - As above but during LCP (default 100M)

      Depricated old NoOfDiskPagesToDisk*

    - Change NoOfFragmentLogFiles default to 16 (1Gb)


storage/ndb/include/kernel/signaldata/BackupContinueB.hpp:
  Add possibility to limitat of disk write speed in backup
storage/ndb/include/mgmapi/mgmapi_config_parameters.h:
  Add possibility to limitat of disk write speed in backup
storage/ndb/src/kernel/blocks/backup/Backup.cpp:
  Add possibility to limitat of disk write speed in backup
storage/ndb/src/kernel/blocks/backup/Backup.hpp:
  Add possibility to limitat of disk write speed in backup
storage/ndb/src/kernel/blocks/backup/BackupInit.cpp:
  Add possibility to limitat of disk write speed in backup
storage/ndb/src/mgmsrv/ConfigInfo.cpp:
  Add possibility to limitat of disk write speed in backup
  Change NoOfFragmentLogFiles default to 16 (1Gb)
  DiskSyncSize
  DiskCheckpointSpeed
  DiskCheckpointSpeedInRestart
storage/ndb/src/mgmsrv/InitConfigFileParser.cpp:
  Handle deprication warning also in my.cnf format

7 files changed, 242 insertions, 47 deletions

diff --git a/storage/ndb/include/kernel/signaldata/BackupContinueB.hpp b/storage/ndb/include/kernel/signaldata/BackupContinueB.hpp
index fe3f48444ec..9035c6f8140 100644
--- a/storage/ndb/include/kernel/signaldata/BackupContinueB.hpp
+++ b/storage/ndb/include/kernel/signaldata/BackupContinueB.hpp
@@ -32,7 +32,8 @@ private:
     BUFFER_FULL_SCAN  = 2,
     BUFFER_FULL_FRAG_COMPLETE = 3,
     BUFFER_FULL_META  = 4,
-    BACKUP_FRAGMENT_INFO = 5
+    BACKUP_FRAGMENT_INFO = 5,
+    RESET_DISK_SPEED_COUNTER = 6
   };
 };
 
diff --git a/storage/ndb/include/mgmapi/mgmapi_config_parameters.h b/storage/ndb/include/mgmapi/mgmapi_config_parameters.h
index 78d34b31bbb..d1feaa1a7d3 100644
--- a/storage/ndb/include/mgmapi/mgmapi_config_parameters.h
+++ b/storage/ndb/include/mgmapi/mgmapi_config_parameters.h
@@ -92,6 +92,10 @@
 #define CFG_DB_DISK_PAGE_BUFFER_MEMORY 160
 #define CFG_DB_STRING_MEMORY          161
 
+#define CFG_DB_DISK_SYNCH_SIZE        163
+#define CFG_DB_CHECKPOINT_SPEED       164
+#define CFG_DB_CHECKPOINT_SPEED_SR    165
+
 #define CFG_DB_SGA                    198 /* super pool mem */
 #define CFG_DB_DATA_MEM_2             199 /* used in special build in 5.1 */
 
diff --git a/storage/ndb/src/kernel/blocks/backup/Backup.cpp b/storage/ndb/src/kernel/blocks/backup/Backup.cpp
index 2e8d8b548ce..d170b3f5a6a 100644
--- a/storage/ndb/src/kernel/blocks/backup/Backup.cpp
+++ b/storage/ndb/src/kernel/blocks/backup/Backup.cpp
@@ -84,6 +84,16 @@ Backup::execSTTOR(Signal* signal)
   const Uint32 startphase  = signal->theData[1];
   const Uint32 typeOfStart = signal->theData[7];
 
+  if (startphase == 1)
+  {
+    m_curr_disk_write_speed = c_defaults.m_disk_write_speed_sr;
+    m_overflow_disk_write = 0;
+    m_reset_disk_speed_time = NdbTick_CurrentMillisecond();
+    m_reset_delay_used = Backup::DISK_SPEED_CHECK_DELAY;
+    signal->theData[0] = BackupContinueB::RESET_DISK_SPEED_COUNTER;
+    sendSignalWithDelay(BACKUP_REF, GSN_CONTINUEB, signal,
+                        Backup::DISK_SPEED_CHECK_DELAY, 1);
+  }
   if (startphase == 3) {
     jam();
     g_TypeOfStart = typeOfStart;
@@ -92,6 +102,11 @@ Backup::execSTTOR(Signal* signal)
     return;
   }//if
 
+  if (startphase == 7)
+  {
+    m_curr_disk_write_speed = c_defaults.m_disk_write_speed;
+  }
+
   if(startphase == 7 && g_TypeOfStart == NodeState::ST_INITIAL_START &&
      c_masterNodeId == getOwnNodeId()){
     jam();
@@ -170,6 +185,42 @@ Backup::execCONTINUEB(Signal* signal)
   const Uint32 Tdata2 = signal->theData[2];
   
   switch(Tdata0) {
+  case BackupContinueB::RESET_DISK_SPEED_COUNTER:
+  {
+    /*
+      Adjust for upto 10 millisecond delay of this signal. Longer
+      delays will not be handled, in this case the system is most
+      likely under too high load and it won't matter very much that
+      we decrease the speed of checkpoints.
+
+      We use a technique where we allow an overflow write in one
+      period. This overflow will be removed from the next period
+      such that the load will at average be as specified.
+    */
+    int delay_time = m_reset_delay_used;
+    NDB_TICKS curr_time = NdbTick_CurrentMillisecond();
+    int sig_delay = curr_time - m_reset_disk_speed_time;
+
+    m_words_written_this_period = m_overflow_disk_write;
+    m_overflow_disk_write = 0;
+    m_reset_disk_speed_time = curr_time;
+
+    if (sig_delay > delay_time + 10)
+      delay_time = Backup::DISK_SPEED_CHECK_DELAY - 10;
+    else if (sig_delay < delay_time - 10)
+      delay_time = Backup::DISK_SPEED_CHECK_DELAY + 10;
+    else
+      delay_time = Backup::DISK_SPEED_CHECK_DELAY - (sig_delay - delay_time);
+    m_reset_delay_used= delay_time;
+    signal->theData[0] = BackupContinueB::RESET_DISK_SPEED_COUNTER;
+    sendSignalWithDelay(BACKUP_REF, GSN_CONTINUEB, signal, delay_time, 1);
+#if 0
+    ndbout << "Signal delay was = " << sig_delay;
+    ndbout << " Current time = " << curr_time << endl;
+    ndbout << " Delay time will be = " << delay_time << endl << endl;
+#endif
+    break;
+  }
   case BackupContinueB::BACKUP_FRAGMENT_INFO:
   {
     const Uint32 ptr_I = Tdata1;
@@ -202,8 +253,8 @@ Backup::execCONTINUEB(Signal* signal)
     fragInfo->FragmentNo = htonl(fragPtr_I);
     fragInfo->NoOfRecordsLow = htonl(fragPtr.p->noOfRecords & 0xFFFFFFFF);
     fragInfo->NoOfRecordsHigh = htonl(fragPtr.p->noOfRecords >> 32);
-    fragInfo->FilePosLow = htonl(0 & 0xFFFFFFFF);
-    fragInfo->FilePosHigh = htonl(0 >> 32);
+    fragInfo->FilePosLow = htonl(0);
+    fragInfo->FilePosHigh = htonl(0);
 
     filePtr.p->operation.dataBuffer.updateWritePtr(sz);
 
@@ -938,7 +989,7 @@ Backup::execBACKUP_REQ(Signal* signal)
     return;
   }//if
 
-  if (m_diskless)
+  if (c_defaults.m_diskless)
   {
     sendBackupRef(senderRef, flags, signal, senderData, 
 		  BackupRef::CannotBackupDiskless);
@@ -2610,9 +2661,10 @@ Backup::openFiles(Signal* signal, BackupRecordPtr ptr)
     FsOpenReq::OM_WRITEONLY | 
     FsOpenReq::OM_TRUNCATE |
     FsOpenReq::OM_CREATE | 
-    FsOpenReq::OM_APPEND;
+    FsOpenReq::OM_APPEND |
+    FsOpenReq::OM_AUTOSYNC;
   FsOpenReq::v2_setCount(req->fileNumber, 0xFFFFFFFF);
-  
+  req->auto_sync_size = c_defaults.m_disk_synch_size;
   /**
    * Ctl file
    */
@@ -3881,6 +3933,69 @@ Backup::execFSAPPENDCONF(Signal* signal)
   checkFile(signal, filePtr);
 }
 
+/*
+  This routine handles two problems with writing to disk during local
+  checkpoints and backups. The first problem is that we need to limit
+  the writing to ensure that we don't use too much CPU and disk resources
+  for backups and checkpoints. The perfect solution to this is to use
+  a dynamic algorithm that adapts to the environment. Until we have
+  implemented this we can satisfy ourselves with an algorithm that
+  uses a configurable limit.
+
+  The second problem is that in Linux we can get severe problems if we
+  write very much to the disk without synching. In the worst case we
+  can have Gigabytes of data in the Linux page cache before we reach
+  the limit of how much we can write. If this happens the performance
+  will drop significantly when we reach this limit since the Linux flush
+  daemon will spend a few minutes on writing out the page cache to disk.
+  To avoid this we ensure that a file never have more than a certain
+  amount of data outstanding before synch. This variable is also
+  configurable.
+*/
+bool
+Backup::ready_to_write(bool ready, Uint32 sz, bool eof, BackupFile *fileP)
+{
+#if 0
+  ndbout << "ready_to_write: ready = " << ready << " eof = " << eof;
+  ndbout << " sz = " << sz << endl;
+  ndbout << "words this period = " << m_words_written_this_period;
+  ndbout << endl << "overflow disk write = " << m_overflow_disk_write;
+  ndbout << endl << "Current Millisecond is = ";
+  ndbout << NdbTick_CurrentMillisecond() << endl;
+#endif
+  if ((ready || eof) &&
+      m_words_written_this_period <= m_curr_disk_write_speed)
+  {
+    /*
+      We have a buffer ready to write or we have reached end of
+      file and thus we must write the last before closing the
+      file.
+      We have already check that we are allowed to write at this
+      moment. We only worry about history of last 100 milliseconds.
+      What happened before that is of no interest since a disk
+      write that was issued more than 100 milliseconds should be
+      completed by now.
+    */
+    int overflow;
+    m_words_written_this_period += sz;
+    overflow = m_words_written_this_period - m_curr_disk_write_speed;
+    if (overflow > 0)
+      m_overflow_disk_write = overflow;
+#if 0
+    ndbout << "Will write with " << endl;
+    ndbout << endl;
+#endif
+    return true;
+  }
+  else
+  {
+#if 0
+    ndbout << "Will not write now" << endl << endl;
+#endif
+    return false;
+  }
+}
+
 void
 Backup::checkFile(Signal* signal, BackupFilePtr filePtr)
 {
@@ -3890,35 +4005,23 @@ Backup::checkFile(Signal* signal, BackupFilePtr filePtr)
 #endif
 
   OperationRecord & op = filePtr.p->operation;
-  
-  Uint32 * tmp, sz; bool eof;
-  if(op.dataBuffer.getReadPtr(&tmp, &sz, &eof)) 
+  Uint32 *tmp = NULL;
+  Uint32 sz = 0;
+  bool eof = FALSE;
+  bool ready = op.dataBuffer.getReadPtr(&tmp, &sz, &eof); 
+#if 0
+  ndbout << "Ptr to data = " << hex << tmp << endl;
+#endif
+  if (!ready_to_write(ready, sz, eof, filePtr.p))
   {
     jam();
-    
-    jam();
-    FsAppendReq * req = (FsAppendReq *)signal->getDataPtrSend();
-    req->filePointer   = filePtr.p->filePointer;
-    req->userPointer   = filePtr.i;
-    req->userReference = reference();
-    req->varIndex      = 0;
-    req->offset        = tmp - c_startOfPages;
-    req->size          = sz;
-    
-    sendSignal(NDBFS_REF, GSN_FSAPPENDREQ, signal, 
-	       FsAppendReq::SignalLength, JBA);
-    return;
-  }
-  
-  if(!eof) {
-    jam();
     signal->theData[0] = BackupContinueB::BUFFER_UNDERFLOW;
     signal->theData[1] = filePtr.i;
-    sendSignalWithDelay(BACKUP_REF, GSN_CONTINUEB, signal, 50, 2);
+    sendSignalWithDelay(BACKUP_REF, GSN_CONTINUEB, signal, 20, 2);
     return;
-  }//if
-  
-  if(sz > 0) {
+  }
+  else if (sz > 0)
+  {
     jam();
     FsAppendReq * req = (FsAppendReq *)signal->getDataPtrSend();
     req->filePointer   = filePtr.p->filePointer;
@@ -3926,13 +4029,14 @@ Backup::checkFile(Signal* signal, BackupFilePtr filePtr)
     req->userReference = reference();
     req->varIndex      = 0;
     req->offset        = tmp - c_startOfPages;
-    req->size          = sz; // Round up
+    req->size          = sz;
+    req->synch_flag    = 0;
     
     sendSignal(NDBFS_REF, GSN_FSAPPENDREQ, signal, 
 	       FsAppendReq::SignalLength, JBA);
     return;
-  }//if
-
+  }
+  
 #ifdef DEBUG_ABORT
   Uint32 running= filePtr.p->fileRunning;
   Uint32 closing= filePtr.p->fileClosing;
@@ -4214,16 +4318,15 @@ Backup::closeFiles(Signal* sig, BackupRecordPtr ptr)
       continue;
     }//if
 
+    filePtr.p->operation.dataBuffer.eof();
     if(filePtr.p->fileRunning == 1){
       jam();
 #ifdef DEBUG_ABORT
       ndbout_c("Close files fileRunning == 1, filePtr.i=%u", filePtr.i);
 #endif
-      filePtr.p->operation.dataBuffer.eof();
     } else {
       jam();
       filePtr.p->fileClosing = 1;
-      filePtr.p->operation.dataBuffer.eof();
       checkFile(sig, filePtr); // make sure we write everything before closing
 
       FsCloseReq * req = (FsCloseReq *)sig->getDataPtrSend();
@@ -4712,8 +4815,10 @@ Backup::lcp_open_file(Signal* signal, BackupRecordPtr ptr)
     FsOpenReq::OM_WRITEONLY | 
     FsOpenReq::OM_TRUNCATE |
     FsOpenReq::OM_CREATE | 
-    FsOpenReq::OM_APPEND;
+    FsOpenReq::OM_APPEND |
+    FsOpenReq::OM_AUTOSYNC;
   FsOpenReq::v2_setCount(req->fileNumber, 0xFFFFFFFF);
+  req->auto_sync_size = c_defaults.m_disk_synch_size;
   
   TablePtr tabPtr;
   FragmentPtr fragPtr;
diff --git a/storage/ndb/src/kernel/blocks/backup/Backup.hpp b/storage/ndb/src/kernel/blocks/backup/Backup.hpp
index afacf01ab2f..4f54918ed9d 100644
--- a/storage/ndb/src/kernel/blocks/backup/Backup.hpp
+++ b/storage/ndb/src/kernel/blocks/backup/Backup.hpp
@@ -33,6 +33,7 @@
 #include <blocks/mutexes.hpp>
 
 #include <NdbTCP.h>
+#include <NdbTick.h>
 #include <Array.hpp>
 
 /**
@@ -522,6 +523,11 @@ public:
     Uint32 m_minWriteSize;
     Uint32 m_maxWriteSize;
     Uint32 m_lcp_buffer_size;
+    
+    Uint32 m_disk_write_speed_sr;
+    Uint32 m_disk_write_speed;
+    Uint32 m_disk_synch_size;
+    Uint32 m_diskless;
   };
   
   /**
@@ -533,8 +539,17 @@ public:
   NdbNodeBitmask c_aliveNodes;
   DLList<BackupRecord> c_backups;
   Config c_defaults;
-  Uint32 m_diskless;
 
+  /*
+    Variables that control checkpoint to disk speed
+  */
+  Uint32 m_curr_disk_write_speed;
+  Uint32 m_words_written_this_period;
+  Uint32 m_overflow_disk_write;
+  Uint32 m_reset_delay_used;
+  NDB_TICKS m_reset_disk_speed_time;
+  static const int  DISK_SPEED_CHECK_DELAY = 100;
+  
   STATIC_CONST(NO_OF_PAGES_META_FILE = MAX_WORDS_META_FILE/BACKUP_WORDS_PER_PAGE);
 
   /**
@@ -631,6 +646,8 @@ public:
   void lcp_open_file_done(Signal*, BackupRecordPtr);
   void lcp_close_file_conf(Signal* signal, BackupRecordPtr);
   void lcp_send_end_lcp_conf(Signal* signal, BackupRecordPtr);
+
+  bool ready_to_write(bool ready, Uint32 sz, bool eof, BackupFile *fileP);
 };
 
 inline
diff --git a/storage/ndb/src/kernel/blocks/backup/BackupInit.cpp b/storage/ndb/src/kernel/blocks/backup/BackupInit.cpp
index 38a60ac04d6..4cbe0c32a29 100644
--- a/storage/ndb/src/kernel/blocks/backup/BackupInit.cpp
+++ b/storage/ndb/src/kernel/blocks/backup/BackupInit.cpp
@@ -146,8 +146,28 @@ Backup::execREAD_CONFIG_REQ(Signal* signal)
     m_ctx.m_config.getOwnConfigIterator();
   ndbrequire(p != 0);
 
+  c_defaults.m_disk_write_speed = 10 * (1024 * 1024);
+  c_defaults.m_disk_write_speed_sr = 100 * (1024 * 1024);
+  c_defaults.m_disk_synch_size = 4 * (1024 * 1024);
+  
   Uint32 noBackups = 0, noTables = 0, noAttribs = 0, noFrags = 0;
-  ndbrequire(!ndb_mgm_get_int_parameter(p, CFG_DB_DISCLESS, &m_diskless));
+  ndbrequire(!ndb_mgm_get_int_parameter(p, CFG_DB_DISCLESS, 
+					&c_defaults.m_diskless));
+  ndb_mgm_get_int_parameter(p, CFG_DB_CHECKPOINT_SPEED_SR,
+			    &c_defaults.m_disk_write_speed_sr);
+  ndb_mgm_get_int_parameter(p, CFG_DB_CHECKPOINT_SPEED,
+			    &c_defaults.m_disk_write_speed);
+  ndb_mgm_get_int_parameter(p, CFG_DB_DISK_SYNCH_SIZE,
+			    &c_defaults.m_disk_synch_size);
+
+  /*
+    We adjust the disk speed parameters from bytes per second to rather be
+    words per 100 milliseconds. We convert disk synch size from bytes per
+    second to words per second.
+  */
+  c_defaults.m_disk_write_speed /= (4 * 10);
+  c_defaults.m_disk_write_speed_sr /= (4 * 10);
+
   ndb_mgm_get_int_parameter(p, CFG_DB_PARALLEL_BACKUPS, &noBackups);
   //  ndbrequire(!ndb_mgm_get_int_parameter(p, CFG_DB_NO_TABLES, &noTables));
   ndbrequire(!ndb_mgm_get_int_parameter(p, CFG_DICT_TABLE, &noTables));
diff --git a/storage/ndb/src/mgmsrv/ConfigInfo.cpp b/storage/ndb/src/mgmsrv/ConfigInfo.cpp
index 6c172a29819..fb15e35ecc7 100644
--- a/storage/ndb/src/mgmsrv/ConfigInfo.cpp
+++ b/storage/ndb/src/mgmsrv/ConfigInfo.cpp
@@ -877,7 +877,7 @@ const ConfigInfo::ParamInfo ConfigInfo::m_ParamInfo[] = {
     ConfigInfo::CI_USED,
     false,
     ConfigInfo::CI_INT,
-    "8",
+    "16",
     "3",
     STR_VALUE(MAX_INT_RNIL) },
 
@@ -952,8 +952,8 @@ const ConfigInfo::ParamInfo ConfigInfo::m_ParamInfo[] = {
     CFG_DB_LCP_DISC_PAGES_TUP_SR,
     "NoOfDiskPagesToDiskDuringRestartTUP",
     DB_TOKEN,
-    "?",
-    ConfigInfo::CI_USED,
+    "DiskCheckpointSpeedSr",
+    ConfigInfo::CI_DEPRICATED,
     true,
     ConfigInfo::CI_INT,
     "40",
@@ -964,8 +964,8 @@ const ConfigInfo::ParamInfo ConfigInfo::m_ParamInfo[] = {
     CFG_DB_LCP_DISC_PAGES_TUP,
     "NoOfDiskPagesToDiskAfterRestartTUP",
     DB_TOKEN,
-    "?",
-    ConfigInfo::CI_USED,
+    "DiskCheckpointSpeed",
+    ConfigInfo::CI_DEPRICATED,
     true,
     ConfigInfo::CI_INT,
     "40",
@@ -976,8 +976,8 @@ const ConfigInfo::ParamInfo ConfigInfo::m_ParamInfo[] = {
     CFG_DB_LCP_DISC_PAGES_ACC_SR,
     "NoOfDiskPagesToDiskDuringRestartACC",
     DB_TOKEN,
-    "?",
-    ConfigInfo::CI_USED,
+    "DiskCheckpointSpeedSr",
+    ConfigInfo::CI_DEPRICATED,
     true,
     ConfigInfo::CI_INT,
     "20",
@@ -988,8 +988,8 @@ const ConfigInfo::ParamInfo ConfigInfo::m_ParamInfo[] = {
     CFG_DB_LCP_DISC_PAGES_ACC,
     "NoOfDiskPagesToDiskAfterRestartACC",
     DB_TOKEN,
-    "?",
-    ConfigInfo::CI_USED,
+    "DiskCheckpointSpeed",
+    ConfigInfo::CI_DEPRICATED,
     true,
     ConfigInfo::CI_INT,
     "20",
@@ -1192,6 +1192,42 @@ const ConfigInfo::ParamInfo ConfigInfo::m_ParamInfo[] = {
     0, 0 },
   
   { 
+    CFG_DB_DISK_SYNCH_SIZE,
+    "DiskSyncSize",
+    DB_TOKEN,
+    "Data written to a file before a synch is forced",
+    ConfigInfo::CI_USED,
+    false,
+    ConfigInfo::CI_INT,
+    "4M",
+    "32k",
+    STR_VALUE(MAX_INT_RNIL) },
+  
+  { 
+    CFG_DB_CHECKPOINT_SPEED,
+    "DiskCheckpointSpeed",
+    DB_TOKEN,
+    "Bytes per second allowed to be written by checkpoint",
+    ConfigInfo::CI_USED,
+    false,
+    ConfigInfo::CI_INT,
+    "10M",
+    "1M",
+    STR_VALUE(MAX_INT_RNIL) },
+  
+  { 
+    CFG_DB_CHECKPOINT_SPEED_SR,
+    "DiskCheckpointSpeedInRestart",
+    DB_TOKEN,
+    "Bytes per second allowed to be written by checkpoint during restart",
+    ConfigInfo::CI_USED,
+    false,
+    ConfigInfo::CI_INT,
+    "100M",
+    "1M",
+    STR_VALUE(MAX_INT_RNIL) },
+  
+  { 
     CFG_DB_BACKUP_MEM,
     "BackupMemory",
     DB_TOKEN,
diff --git a/storage/ndb/src/mgmsrv/InitConfigFileParser.cpp b/storage/ndb/src/mgmsrv/InitConfigFileParser.cpp
index 68a5f02f4c5..bf5cb9d726e 100644
--- a/storage/ndb/src/mgmsrv/InitConfigFileParser.cpp
+++ b/storage/ndb/src/mgmsrv/InitConfigFileParser.cpp
@@ -655,6 +655,18 @@ InitConfigFileParser::store_in_properties(Vector<struct my_option>& options,
 			m_info->getMax(ctx.m_currentInfo, fname));
 	return false;
       }
+
+      ConfigInfo::Status status = m_info->getStatus(ctx.m_currentInfo, fname);
+      if (status == ConfigInfo::CI_DEPRICATED) {
+	const char * desc = m_info->getDescription(ctx.m_currentInfo, fname);
+	if(desc && desc[0]){
+	  ctx.reportWarning("[%s] %s is depricated, use %s instead", 
+			    ctx.fname, fname, desc);
+	} else if (desc == 0){
+	  ctx.reportWarning("[%s] %s is depricated", ctx.fname, fname);
+	} 
+      }
+      
       if (options[i].var_type == GET_INT)
 	ctx.m_currentSection->put(options[i].name, (Uint32)value_int);
       else