Charm++ parallel programming system

[Ankur Narang <annarang AT in.ibm.com>]

Hi Yanhua,

 Thanks for the details below.
I linked using -balance workstealing. I ran using $T3WL +cs +LBDebug 3,
but I don't see the per processor nodes processed/generated information.
Is a separate file generated for this information. Here is a cutout from
the output of Charm++ run:

##################
Charm++> work stealing seed load
balancer
CharmLB> Verbose level 3, load balancing
period: 0.5 seconds
CharmLB< Toplogy torus_nd_5 alpha:
3.50000e-05s beta: 8.5e--09s
stats unavailable in optimized version.
Ignoring.....
#####################


Is it because of this last msg that
it is an optimized version ??

Thanks,
Regards,
Ankur.

Senior Researcher (RSM)
Research Lead - HPC Analytics
IBM Research India
4, Block C, Institutional Area,
Phase II, Vasant Kunj,
New Delhi - 110070




From:      
 Yanhua Sun <sun51 AT illinois.edu>
To:      
 Abhinav Bhatele <bhatele AT illinoisalumni.org>,

Cc:      
 Ankur Narang/India/IBM@IBMIN,
Charm Mailing List <charm AT cs.illinois.edu>
Date:      
 07/08/2013 08:07 PM
Subject:    
   Re: Fwd: Charm++
Qs
Sent by:    
   yanh.sun AT gmail.com

---

Hi Ankur,

Before I tell you how, I would like to explain a little
bit about  Charm++ load balancer. 
In Charm++, there are two types of load balancer based
on the application.
For the applications which has persistence features (mostly
the iterative applications),
we have the persistent load balancer. It takes the history
information as prediction of load balancing
strategy. This load balancer is most known in Charm.  This
is also what you tried. 

HOWEVER, for applications without persistence feature,
that means the history can not be used to predict future, there is no iterative
steps.  We have Seed load balancer. This is what we used in these
state space search applications, including UTS.  The details can be
found here http://charm.cs.illinois.edu/manuals/html/charm++/7.html#SECTION01670000000000000000

In order to print the load balancing information, you
can use +cs, which prints out how many tree nodes are generated processed
on each processor. 

Let me know whether this works. 

Best,

Yanhua





 


On Mon, Jul 8, 2013 at 1:51 AM, Abhinav Bhatele <bhatele AT illinoisalumni.org>
wrote:
I am ccing Yanhua who should be able to answer your questions.

- Abhinav


On Sun, Jul 7, 2013 at 9:24 PM, Ankur Narang <annarang AT in.ibm.com>
wrote:
Hi Abhinav, 

  We resolved the Kmeans code issue. So, now would only require help
for UTS. The UTS code is in the Charm++ repository at:
CHARM/charm/examples/charm++/state_space_searchengine/UnbalancedTreeSearch_SE/

I need to print out the load distribution information across the compute
nodes when large UTS binomial trees are run. I compiled the code with -module
CommonLBs and ran the code with +balancer <name> +LBDebug 1. But,
I am not getting any load balance information. Preferably, would like to
have this each time the load balancer runs. I have three specific qs: 
(1) Is the load balancer running by default when give the above compile
and run options ? 
(2) If not then, do we need to use AtSync() ? 
(3) If AtSync() is needed to run the load balancing algorithm then, how
to write that code. I think we need resumeFromSync() as well. If there
exists some code for UTS with such additions for load balancing then that
will be helpful. 

Thanks, 
Ankur. 

Senior Researcher (RSM)
Research Lead - HPC Analytics
IBM Research India
4, Block C, Institutional Area,
Phase II, Vasant Kunj,
New Delhi - 110070




From:      
 Abhinav Bhatele <bhatele AT illinoisalumni.org>
To:        Ankur
Narang/India/IBM@IBMIN, 
Date:        07/08/2013
09:36 AM 
Subject:        Re:
Fwd: Charm++ Qs 
Sent by:    
   bhatele AT gmail.com

---

Can you give me some more details? what is the UTS code and is the kmeans
code in the charm repository? 


On Sun, Jul 7, 2013 at 8:50 PM, Ankur Narang <annarang AT in.ibm.com>
wrote: 
Sure, Thanks Abhinav, 

  If you know some person in particular in Charm++ group who could
help, that will be great, 
Ankur. 

Senior Researcher (RSM)
Research Lead - HPC Analytics
IBM Research India
4, Block C, Institutional Area,
Phase II, Vasant Kunj,
New Delhi - 110070




From:      
 Abhinav Bhatele <bhatele AT illinoisalumni.org>
To:        Charm
Mailing List <charm AT cs.illinois.edu>,
Cc:        Ankur
Narang/India/IBM@IBMIN 
Date:        07/06/2013
06:58 AM 
Subject:        Fwd:
Charm++ Qs 
Sent by:        bhatele AT gmail.com

---

Hi Ankur,

Sorry for the delay in response. I am forwarding your e-mail to the Charm
mailing list and someone should be able to answer your questions there.

- Abhinav


Begin forwarded message: 

From: Ankur Narang <annarang AT in.ibm.com>
Subject: Charm++ Qs 
Date: June 21, 2013 4:29:02 AM PDT 
To: <bhatele AT llnl.gov>

Hi Abhinav, 
                     
  Hope you are doing great. I had couple of qs on Charm++

1) We would like to measure the load balance across the compute nodes during
the run of UTS on Charm++. How do we do that using the available code of
UTS in the Charm++ distribution ? 

2) We have a kmeans code in Charm++ that does not scale with increasing
nodes. How do we resolve this issue ? 

Also, my team is hiring HPC professionals. If you know some folks then
let me know, 

Thanks, 
Ankur. 

Senior Researcher (RSM)
Research Lead - HPC Analytics
IBM Research India
4, Block C, Institutional Area,
Phase II, Vasant Kunj,
New Delhi - 110070 


-- 
Abhinav Bhatele, people.llnl.gov/bhatele
Center for Applied Scientific Computing, Lawrence Livermore National
Laboratory 



-- 
Abhinav Bhatele, people.llnl.gov/bhatele
Center for Applied Scientific Computing, Lawrence Livermore National
Laboratory 



-- 
Abhinav Bhatele, people.llnl.gov/bhatele
Center for Applied Scientific Computing, Lawrence Livermore National
Laboratory 



-- 
Yours Sincerely,
Yanhua Sun 
University of Illinois at Urbana-Champaign,