[mpich-discuss] Hydra binding questions

[Jain, Rohit]

Pavan,

Map info (sent in earlier mail) was from a machine with hyper-threading enabled. So, it had 8 physical core and 16 virtual cores. Is binding done on virtual cores as well?


Here is mapping info from a machine with 12 physical cores. Cores/sockets binding still mapping only 8 cores, which looks like a bug?

topomap: 
(MACHINE,
(NODE,mem_size:34357497856,
(SOCKET,
(CORE,(THREAD,pid:0)),
(CORE,(THREAD,pid:2)),
(CORE,(THREAD,pid:4)),
(CORE,(THREAD,pid:6)),
(CORE,(THREAD,pid:8)),
(CORE,(THREAD,pid:10)))),
(NODE,mem_size:34359730176,
(SOCKET,
(CORE,(THREAD,pid:1)),
(CORE,(THREAD,pid:3)),
(CORE,(THREAD,pid:5)),
(CORE,(THREAD,pid:7)),
(CORE,(THREAD,pid:9)),
(CORE,(THREAD,pid:11)))))

Cpu:cores:  processmap: (0;2;4;6;;;1;3;5;7;;)
Cou:sockets: processmap: (0;1;2;3;4;5;6;7;;;;)
Rr : processmap: (0;1;2;3;4;5;6;7;8;9;10;11)


Regards,
Rohit


-----Original Message-----
From: Pavan Balaji [mailto:balaji at mcs.anl.gov] 
Sent: Wednesday, August 10, 2011 3:35 PM
To: Jain, Rohit
Cc: mpich-discuss at mcs.anl.gov
Subject: Re: [mpich-discuss] Hydra binding questions

Hi Rohit,

On 08/10/2011 02:56 PM, Jain, Rohit wrote:
> Okay. Let me know when you have it ready and few instructions on how
> to get it from trunk and use it with 1.4.

The changes have been committed to trunk as of r8871. You can download 
the latest nightly snapshot of Hydra from 
http://www.mcs.anl.gov/research/projects/mpich2/downloads/tarballs/nightly/hydra/ 
and build it. It'll give you an mpiexec (and some other utility 
executables), that you can use to run your program. You can still 
compile your application with mpicc and friends from the 1.4 
installation; you just need to use the new mpiexec to run it.

When you pass the -verbose flag, it'll show you both the hardware 
topology (labeled as 'topomap') and the process binding information 
('processmap'). There might be too much additional information in the 
debug output, so you might want to grep for these flags in the output.

  -- Pavan

-- 
Pavan Balaji
http://www.mcs.anl.gov/~balaji