mtmaps (OpenMP)

[Woweur]

Uptime ? ~3 hours, online players ? ~3000.

But you need include OMP, and i don't know if it's work with ACE mtmaps, but i can create a little patch for ACE if you want.

I have just add that :

   #pragma omp critical(spellLock)
   {
       m_Events.Update( p_time );
       _UpdateSpells( p_time );
   }

[Span1sh]

Uptime ? ~3 hours, online players ? ~3000.

But you need include OMP, and i don't know if it's work with ACE mtmaps, but i can create a little patch for ACE if you want.

I have just add that :

   #pragma omp critical(spellLock)
   {
       m_Events.Update( p_time );
       _UpdateSpells( p_time );
   }

Yes pls create that patch. And I am very interested why you have chosen OpenMP against ACE? Is it much more stable?

P/S What core rev do you use? Is there any custom patches except MTMAPS?

Best Regards

Thx a lot

[Woweur]

Just because it's easy code than ACE =P

My rev ? a old rev with some custom modification (by me) for a spell, stability, instance ...etc

I'm go to write a little patch for you

[Woweur]

lock for spell update with ACE multithread map : http://paste2.org/p/688715

[Span1sh]

lock for spell update with ACE multithread map : http://paste2.org/p/688715

Thx. I'll try that and report...

[rusfighter]

Im testing now it. Btw i have a question about CPU usage , if i use command "top" then the usage is about 50% for the PID of the core , but the average is 7% , so it means that i use 7% of the CPU and take alot of more and i can cross 100% for the PID?

I have quad core

BTW . homuch do i need to set ACEmap thread? i have set it at 1

[Woweur]

Personaly, i set 16 threads because my server has 16 core

For show % of use by thread, i use top command and i press 1 on putty.

[Unik2psc]

commit 9ae413e4a5568c8ccfbf0cdf26510675fa352126

Author: izidor.millenium <[email protected]>

Date: Thu Feb 25 18:26:32 2010 +0100

- ACE multithread map lock for spell update

Signed-off-by: izidor.millenium <[email protected]>

diff --git a/src/game/Unit.cpp b/src/game/Unit.cpp

index cbe69fd..1e1a2e9 100644

--- a/src/game/Unit.cpp

+++ b/src/game/Unit.cpp

@@ -310,8 +310,10 @@ void Unit::Update( uint32 p_time )

// WARNING! Order of execution here is important, do not change.

// Spells must be processed with event system BEFORE they go to _UpdateSpells.

// Or else we may have some SPELL_STATE_FINISHED spells stalled in pointers, that is bad.

+ sWorld.m_spellUpdateLock.acquire();

m_Events.Update( p_time );

_UpdateSpells( p_time );

+ sWorld.m_spellUpdateLock.release();

CleanupDeletedAuras();

diff --git a/src/game/World.h b/src/game/World.h

index 8a4b370..961c5bd 100644

--- a/src/game/World.h

+++ b/src/game/World.h

@@ -545,6 +545,7 @@ class World

//used Script version

void SetScriptsVersion(char const* version) { m_ScriptsVersion = version ? version : "unknown scripting library"; }

char const* GetScriptsVersion() { return m_ScriptsVersion.c_str(); }

+ ACE_Thread_Mutex m_spellUpdateLock;

protected:

void _UpdateGameTime();

It is wrong solution.

+ ACE_Thread_Mutex m_spellUpdateLock;

How much is a instance class World ? = 1.

Then we take the mutex. How many units? - Each player, vehicle, npc.

Each update get mutex, in that time others threads is waiting for the queue. This creates a "bottleneck."

How many maps is the npc? - 1.

Why update was to be made from more than one thread?

Solutions must be sought - mutexs are not needed.

If two threads at once the update (), the problem lies elsewhere. Mutexs do not fix it.

(Sorry for my english)

[Woweur]

Yes, I know a general lock isn't a better solution for perfomance, but I haven't a better for now and, with 3000 players online I havnt lag and my uptime is ~15 hours so if you have a better solution post a patch

[Derex]

Woweur mutex patch is perfectly fine, its a common technique to use global mutexes for protecting non-reentrant code. A better solution would be to use #pragma omp critical section than ACE mutexes, cause #pragma omp critical will work better with rest of omp.

As long for openmp, the problem with it is that on some systems the gcc implementation of openmp specification ( libgomp ) is configured to use linux specific optimizations, that cause high cpu ussage ( busy waiting ) and better performance ofc. But having all CPU on 100% is not good if you use the server for other things except mangos. I think the linux specific optimizations appeared in latest gcc releases, so older systems like the stable debian are still using the posix variant which is a bit slower, but doesnt waste that much cpu.

[higi]

anyone got working openmp with gcc 4.4+ ? with gcc 4.4 all threads are on ~100% CPU

and maybe anyone know how to downgrade gcc to 4.2 ubuntu

[crashuncle]

Bad written, delete me.

[DasBlub]

anyone got working openmp with gcc 4.4+ ? with gcc 4.4 all threads are on ~100% CPU

and maybe anyone know how to downgrade gcc to 4.2 ubuntu

you can do this, to get gcc 4.2:

apt-get install gcc-4.2

and then to compile with gcc 4.2 you have to set the variable CC:

export CC=gcc-4.2
make -j4

i'm not sure if you already have to set the variable before configure... you'll need to test it.

-- DasBlub

[seirge]

The problem with high (100%) load with omp can be solved via environment variable. Try the following:

export OMP_WAIT_POLICY=PASSIVE

before starting mangosd

[bodeguero]

Does anyone have any solution to the error concurrent_vector ??

/home/bode/source/mangos/objdir/src/game/../../../src/game/../../dep/tbb/include/tbb/concurrent_vector.h:695: undefined reference to `tbb::internal::concurrent_vector_base_v3::internal_capacity() const'

/home/bode/source/mangos/objdir/src/game/../../../src/game/../../dep/tbb/include/tbb/concurrent_vector.h:695: undefined reference to `tbb::internal::concurrent_vector_base_v3::internal_capacity() const'

collect2: ld returned 1 exit status

make[3]: *** [mangos-worldd] Error 1

[Wojta]

I still geting these crashes, uptime max 10 min: http://pastebin.com/chXktxyt backtrace is always same.

possible that its caused by some SD2 script, anyone has idea what can cause this?

EDIT: Rev 9549, many custom patches, only two related are Visibility and relocation patch by SilverIce and Ambal, second one is MTmaps by Infinity_return

[przemratajczak]

Wojta - probably not related to mtmaps. I am using last patch poseted by infinity + mutex around player and event update and haven't any crash at all from few days (700ppl)

[crashuncle]

So If I get it correct then the best way is to use Infinity's patch with WoWeur's lock for spell update if I want to get the best stability with the 0.12 branch? I don't need the ACE pragma then?

I also got a few conflicts and can't solve two of them one the latest 0.12 branch (397cc2ab0081ef115b27fe5a7b31c3ee2b82e562):

ObjectAccessor.h:

diff a/src/game/ObjectAccessor.h b/src/game/ObjectAccessor.h    (rejected hunks)
@@ -126,14 +131,16 @@ class MANGOS_DLL_DECL ObjectAccessor : public MaNGOS::Singleton<ObjectAccessor,
        void RemoveObject(Player *object) { HashMapHolder<Player>::Remove(object); }

        // TODO: This methods will need lock in MT environment
-        static void LinkMap(Map* map)   { i_mapList.push_back(map); }
-        static void DelinkMap(Map* map) { i_mapList.remove(map); }
+        static void LinkMap(Map* map)   { ACE_Guard<ACE_Thread_Mutex> guard(m_Lock); i_mapList.push_back(map); }
+        static void DelinkMap(Map* map) { ACE_Guard<ACE_Thread_Mutex> guard(m_Lock); i_mapList.remove(map); }
    private:
+        static ACE_Thread_Mutex  m_Lock;
        // TODO: This methods will need lock in MT environment
        // Theoreticaly multiple threads can enter and search in this method but
        // in that case linking/delinking other map should be guarded
        template <class OBJECT> static OBJECT* FindHelper(uint64 guid)
        {
+            ACE_Guard<ACE_Thread_Mutex> guard(m_Lock);
            for (std::list<Map*>::const_iterator i = i_mapList.begin() ; i != i_mapList.end(); ++i)
            {
                if (OBJECT* ret = (*i)->GetObjectsStore().find(guid, (OBJECT*)NULL))

I can't find the related code in the ObjectAccessor.h at all.

Spell.cpp:

    // Lookup target in already in list
-    c
+    for(tbb::concurrent_vector<TargetInfo>::iterator ihit = m_UniqueTargetInfo.begin(); ihit != m_UniqueTargetInfo.end(); ++ihit)

I guess this is also a mistake in the patch and it should be:

     // Lookup target in already in list
-    for(std::list<TargetInfo>::iterator ihit = m_UniqueTargetInfo.begin(); ihit != m_UniqueTargetInfo.end(); ++ihit)
+    for(tbb::concurrent_vector<TargetInfo>::iterator ihit = m_UniqueTargetInfo.begin(); ihit != m_UniqueTargetInfo.end(); ++ihit)

@@ -3377,7 +3441,7 @@ void Spell::WriteSpellGoTargets( WorldPacket * data )
    // m_needAliveTargetMask req for stop channeling if one target die
    uint32 hit  = m_UniqueGOTargetInfo.size(); // Always hits on GO
    uint32 miss = 0;
-    for(std::list<TargetInfo>::iterator ihit = m_UniqueTargetInfo.begin(); ihit != m_UniqueTargetInfo.end(); ++ihit)
+    for(tbb::concurrent_vector<TargetInfo>::iterator ihit = m_UniqueTargetInfo.begin(); ihit != m_UniqueTargetInfo.end(); ++ihit)
    {
        if ((*ihit).effectMask == 0)                  // No effect apply - all immuned add state
        {

I found the function "WriteSpellGoTargets" but it doesn't fits there at all.

@@ -327,7 +327,7 @@ void Spell::EffectSchoolDMG(SpellEffectIndex effect_idx)
                    case 72624: case 72625:                 // Ooze Eruption
                    {
                        uint32 count = 0;
-                        for(std::list<TargetInfo>::iterator ihit= m_UniqueTargetInfo.begin();ihit != m_UniqueTargetInfo.end();++ihit)
+                        for(tbb::concurrent_vector<TargetInfo>::iterator ihit= m_UniqueTargetInfo.begin();ihit != m_UniqueTargetInfo.end();++ihit)
                            if(ihit->effectMask & (1<<effect_idx))
                                ++count;

Only for the Wotlk spell, so quite useless for 0.12 I guess.

[luciolis]

Hello. another problem with a fresh install, sorry.

After the compilation, while packing files :

g++ -g -DMANGOS_DEBUG -DDO_MYSQL -D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64 -O3 -m64 -pipe -msse4.2 -DFD_SETSIZE=4096 -fno-delete-null-pointer-checks -fno-strength-reduce -fno-strict-aliasing -frename-registers -o .libs/mangos-worldd CliRunnable.o Main.o Master.o RASocket.o MaNGOSsoap.o WorldRunnable.o soapC.o soapServer.o ../bindings/ScriptDev2/.libs/libmangosscript.so ../game/libmangosgame.a ../shared/Database/libmangosdatabase.a ../shared/Config/libmangosconfig.a ../shared/Auth/libmangosauth.a ../shared/libmangosshared.a ../shared/vmap/libmangosvmaps.a ../framework/libmangosframework.a ../../dep/src/g3dlite/libg3dlite.a ../../dep/src/gsoap/libgsoap.a -lpthread -Wl,--rpath -Wl,/home/mangos/lib

CliRunnable.o: In function `thread_mutex_lock':

/usr/local/include/ace/OS_NS_Thread.inl:3412: undefined reference to `ACE_OS::mutex_lock(pthread_mutex_t*)'

CliRunnable.o: In function `thread_mutex_unlock':

/usr/local/include/ace/OS_NS_Thread.inl:3491: undefined reference to `ACE_OS::mutex_unlock(pthread_mutex_t*)'

CliRunnable.o: In function `~ACE_Atomic_Op':

/usr/local/include/ace/Atomic_Op_T.h:263: undefined reference to `ACE_Thread_Mutex::~ACE_Thread_Mutex()'

/usr/local/include/ace/Atomic_Op_T.h:263: undefined reference to `ACE_Thread_Mutex::~ACE_Thread_Mutex()'

/usr/local/include/ace/Atomic_Op_T.h:263: undefined reference to `ACE_Thread_Mutex::~ACE_Thread_Mutex()'

Undef reference in all ACE function. It is a problem on my side or a new weird rev. from mangos devs ?

thks !

[yasuno]

I want to put this patch into my core but no matter how i try it or do it i just get the same error "Patch does not have valid e-mail address" anyone know what to do about this? =[

[caeruleaus]

use

git apply

instead of

git am

[yasuno]

So, i used "git apply" and it doesn't report an error, but it doesn't change my files either wtf..

[crashuncle]

Anyone knows how I can solve my patch problems which I wrote in my previous post or got a 0.12 compatible version?

[yasuno]

At least yours will attempt to apply, mine doesn't even try to apply the patch ..

[yasuno]

Does anyone have any solution to the error concurrent_vector ??

/home/bode/source/mangos/objdir/src/game/../../../src/game/../../dep/tbb/include/tbb/concurrent_vector.h:695: undefined reference to `tbb::internal::concurrent_vector_base_v3::internal_capacity() const'

/home/bode/source/mangos/objdir/src/game/../../../src/game/../../dep/tbb/include/tbb/concurrent_vector.h:695: undefined reference to `tbb::internal::concurrent_vector_base_v3::internal_capacity() const'

collect2: ld returned 1 exit status

make[3]: *** [mangos-worldd] Error 1

I have this issue to sadly =[