Charm++ parallel programming system

[Christian Perez <christian.perez AT inria.fr>]

On 05/25/2010 07:22 PM, Gengbin Zheng wrote:
> Is the caller of connect() a thread (I mean is connect() a threaded
> entry function)?
>    
It is not a threaded charm method: when I tried to turn it into a 
[threaded] entry
method it I've got an error when invoking JNI method from this chare. I 
shall investigate it later.
> another way of doing this without using thread is to send component_B
> proxy to component_A, and let component_A directly send its proxy to
> component_B.
>    
That is the alternative solution: the good news is that it works(*),
the not-so-good news is that it make the mapping between components
and charm objects a litte more complex.

(*) I need to use a hack that seems weird to me:

Chare component_A : SimpleInterface {
  void set_s(CProxy_SetSimpleInterface& pssi, CProxy_SimpleInterface& 
psi) {
     pssi.set_si(psi);
  }

works while the following version does not work (psi is changed to 
thishandle)


Chare component_A : SimpleInterface {
  void set_s(CProxy_SetSimpleInterface& pssi, CProxy_SimpleInterface& 
psi) {
     pssi.set_si(thishandle);
  }

psi is obtained from a call to CProxy_component_A::ckNew.

Question: how can I create a Proxy from within a chare?

I tried

     pssi.set_si(CProxy_SimpleInterface(thishandle));

but it fails with the same error :

[0] Assertion "n<len" failed in file cklists.h line 221.
------------- Processor 0 Exiting: Called CmiAbort ------------
Reason:
[0] Stack Traceback:
  [0:0] CmiAbort+0x89  [0x506ff2]
  [0:1] __cmi_assert+0x4b  [0x514a0c]
  [0:2] _ZN5CkVecIPvEixEm+0x36  [0x49c7a8]
  [0:3] /opt/usr/stow/ULCMi/libexec/Charm-launcher [0x4966f9]
  [0:4] _Z15_processHandlerPvP11CkCoreState+0x30d  [0x49712e]
  [0:5] CmiHandleMessage+0x84  [0x50f5f1]
  [0:6] CsdSchedulePoll+0x96  [0x50fb5f]
  [0:7] CsdScheduler+0x23  [0x50f8bf]
  [0:8] CthStandinCode+0xe  [0x50fc5a]
  [0:9] CthStartThread+0x59  [0x4830b7]
  [0:10] /lib/libc.so.6 [0x7f98d9aaf370]
Fatal error on PE 0>

Thanks for your help!

I am still open for any suggestion for making the first version (with 
suspend) working.

Christian

> Gengbin
>
> On Tue, May 25, 2010 at 11:02 AM, Christian Perez
> <christian.perez AT inria.fr>
>   wrote:
>    
> > Ok ... here are the ugly details:
> >
> > We have a JAVA code that interprets an assembly of components, which in this
> > case are implemented in Charm++.
> > The mainchare creates a second chare (called RT) that launched the JVM in an
> > entry method start, which is
> > not threaded (I have problem with the JVM if it is not in the main thread
> > :<)
> >
> > The JAVA code needs to create/connect/execute components. Hence, the java
> > code invokes JNI methods  to call C++ code.
> > The creation turns to be easy (it is a simple call to ckNew).
> > My current problem is with the 'connect'.
> >
> > Logically, the connection operation should be as simple as:
> >      type_of_P p = component_a->get_ref_of_port_P()
> >      component_b->set_ref_of_port_U(p)
> >
> > My current version is in term of ci file is:
> >
> >      chare type_of_P { ... }
> >
> >      chare Component_A : type_of_P {
> >          entry void get_ref_of_port_P(const CkCallbackResumeThread&  cb) {
> > m_si * m = new m_si(); m->ckid=thishandle; } // see def of m_si below
> >      }
> >
> >      chare Component_B {
> >          entry [sync] void set_ref_of_port_U(CkChareID&  ckid);
> >
> > I am using a message to send back a reference to something of type_of_P
> >
> >      message m_si {
> >          CkChareID ckid;
> >     }
> >
> > And the implementation in the jni 'connect' method:
> >
> >       m_si* m;
> >       component_A.get_ref_of_port_P(CkCallbackResumeThread((void*&)m));
> >       component_B.set_ref_of_port_u(m->ckid);
> >
> >
> > Thus, there are in fact two questions:
> > - how to correctly implement the connect method (called by JAVA) ?
> > - how to get&  pass a reference of type 'type_of_P' ? (I got error when
> > CProxy_type_of_P proxy(m->ckid))
> >
> > Does it make sense?
> >
> > Christian
> >
> > On 05/25/2010 05:20 PM, Kale, Laxmikant V wrote:
> >
> > Why not use the sync method solution? (and I assume you declared m1 to be
> > threaded).
> >
> >
> >
> > (And to understand your problem, a little more detail will be useful. Who
> > created the callback? (m1 or m2)? Etc. Showing code fragment omitting all
> > but the relevant calls will be good.)
> >
> >
> >
> > With sync method, it’d look like this
> >
> >
> >
> > M1() { ……; proxy.m2();  … }
> >
> > M2(){… }
> >
> >
> >
> > In .ci: m1 is threaded and m2 is sync.
> >
> >
> >
> > --
> >
> > Laxmikant Kale                  http://charm.cs.uiuc.edu
> >
> > Professor, Computer Science     
> > kale AT illinois.edu
> >
> > 201 N. Goodwin Avenue           Ph:  (217) 244-0094
> >
> > Urbana, IL  61801-2302          FAX: (217) 265-6582
> >
> >
> >
> >
> >
> > From: 
> > charm-bounces AT cs.uiuc.edu
> >  
> > [mailto:charm-bounces AT cs.uiuc.edu]
> >  On Behalf
> > Of Christian Perez
> > Sent: Tuesday, May 25, 2010 10:00 AM
> > To: Filippo Gioachin
> > Cc: 
> > charm AT cs.uiuc.edu
> > Subject: Re: [charm] how to suspend/resume a chare ?
> >
> >
> >
> > Hello,
> >
> > CkCallbackResumeThread() seems to be the perfect answer for one process ...
> > but it fails when using several processes :
> > <<
> > Fatal error on PE 1>  Called thread_delay on different processor than where
> > callback was created
> >      
> > > >          
> > Is there a standard mechanism to solve this issue or should I use an ad-hoc
> > solution by forcing the
> > creation of a local chare?
> >
> > My strong constraint is that m1 can not be spitted in two :(
> >
> > Christian
> >
> > On 05/24/2010 11:01 PM, Filippo Gioachin wrote:
> >
> > Christian,
> >
> >
> >
> > for your purposes, I would suggest looking at the manual on section 3.2.1
> > (Entry Method Attributes) regarding "threaded" and "sync" entry methods. A
> > combination of the two (m1 threaded and m2 sync) should solve your problem.
> >
> > Additionally, callbacks are very useful to resume threads as well; 3.15.2
> > shows how CkCallbackResumeThread can be used to resume m1().
> >
> > Please remember that other entry methods may run while m1 is suspended,
> > including other m1 methods.
> >
> >
> >
> > Filippo
> >
> >
> >
> > On Thu, May 20, 2010 at 03:58, Christian 
> > Perez<christian.perez AT inria.fr>
> > wrote:
> >
> > Hello,
> >
> > I'd like to know how to block a thread. Here is a simple version of what
> > I'd like to do:
> >
> > Chare A:
> >    m1() {
> >      do some computation
> >      invoke method m2 on chare B
> >      wait for the completion of method m2
> >      do something else
> >    }
> >
> > I've tried to use some callback mechanism but the problem is that
> > I can not split m1 in two :(
> >
> > So, is there a wait to suspend the execution (and also how to resume it) ?
> >
> > Thank you!
> >
> > Christian Perez
> >
> > _______________________________________________
> > charm mailing list
> > charm AT cs.uiuc.edu
> > http://lists.cs.uiuc.edu/mailman/listinfo/charm
> >
> >
> > --
> > Filippo Gioachin
> > PhD Candidate
> > Department of Computer Science
> > University of Illinois at Urbana-Champaign
> > Ph: +1-217-333-4764
> >
> >
> >
> > _______________________________________________
> > charm mailing list
> > charm AT cs.uiuc.edu
> > http://lists.cs.uiuc.edu/mailman/listinfo/charm
> >
> >
> >