Performance improvement for persistence

[paul]

Hello,

Here's a patch (sorry, not based on your trunk) that significantly improves the time it takes to write a large number (100K+) objects:

Index: objstrm.cpp
===================================================================
--- objstrm.cpp (revision 77)
+++ objstrm.cpp (revision 78)
@@ -32,6 +32,8 @@
 #if !defined(OWL_EXCEPT_H)
 #include <owl/except.h>
 #endif
+#include <set>
+
 extern void __Trace(LPCSTR lpszFormat, ...);
 inline void __NoTrace(LPCSTR lpszFormat, ...){}
 #if 0
@@ -141,7 +143,7 @@
 {
   //_TRACE("__TStreamableCont::DestroyTypes() Create Types\n");
   // not locking because this functions is called from UnregisterTypes()
-  // and Lock already applyed
+  // and Lock already applied
   //LOCKTYPES 
   delete Types;
   Types = 0;
@@ -207,12 +209,35 @@
   return (int)NPOS;
 }
 //
+/*
 class TSortedTPWObjObjectArray: public TSortedObjectArray<TPWrittenObjects::TPWObj>{
   public:
     TSortedTPWObjObjectArray(int a, int b=0, int c=0): TSortedObjectArray<TPWrittenObjects::TPWObj>(a,b,c){}
     TSortedTPWObjObjectArray(){}
 };
+*/

+// Use a std::set for faster insertion and lookup. The Borland sorted arrays
+// have very poor performance for large numbers of items (O(n^2)).
+class TSortedTPWObjObjectArray
+{
+public:
+   TSortedTPWObjObjectArray(int) {}
+   ~TSortedTPWObjObjectArray() { m_Items.clear(); }
+   void Flush() { m_Items.clear(); }
+   void Add(TPWrittenObjects::TPWObj obj) { m_Items.insert(obj); }
+   P_id_type Find(TPWrittenObjects::TPWObj obj)
+   {
+       std::set<TPWrittenObjects::TPWObj>::const_iterator i = m_Items.find(obj);
+       if(i != m_Items.end()) {
+           return i->Ident;
+       }
+       return 0;
+   }
+private:
+   std::set<TPWrittenObjects::TPWObj> m_Items;
+};
+
 // -----------------------------------------------------------------------------
 const char* TStreamer::StreamableName() const
 {
@@ -345,20 +370,26 @@

 P_id_type TPWrittenObjects::FindObject( TStreamableBase *d )
 {
+  return Data->Find(TPWObj(((_TCHAR*)(void*)d)+1,0));
+/*
   unsigned res = Data->Find(TPWObj(((_TCHAR*)(void*)d)+1,0));
   if (res == NPOS)
     return 0;
   else
    return (*Data)[res].Ident;
+*/
 }

 P_id_type TPWrittenObjects::FindVB( TStreamableBase *d )
 {
+  return Data->Find(TPWObj(d,0));
+/*
   unsigned res = Data->Find(TPWObj(d,0));
   if (res == NPOS)
     return 0;
   else
   return (*Data)[res].Ident;
+*/
 }

 pstream::~pstream()

It reimplemented TSortedTPWObjObjectArray in terms of a std::set instead of the OWL sorted array.

Thanks,
Paul

[Vidar Hasfjord]

Consider creating a ticket for this in Feature Requests (or if you deem the bad performance a bug, in Bugs).

[paul]

Done, see https://sourceforge.net/p/owlnext/feature-requests/116/.

Thanks,
Paul