About training speed with multicore machine

[stevenyslin]

Hello,

I use sphinxtrain to train acoustic model, and I want to speedup the training,
My CPU have 8 core mechine,and my training data have 800 data.

I set the sphinx_train.cfg as follows, and it will take at least 10 minutes,

$CFG_QUEUE_TYPE = "Queue::POSIX";
$CFG_NPART = 8;
$DEC_CFG_NPART = 8; 

But when I using original setting, it will take 3 minutes 12 seconds

$CFG_QUEUE_TYPE = "Queue";
$CFG_NPART = 1;
$DEC_CFG_NPART = 1; 

My question：
Could you tell me what wrong with my settings, thank you so much.

The file as follows is sphinx_train.cfg：
https://drive.google.com/file/d/0BzvKcj5lMO_4M0lxMEEwVjdCcFk/view?usp=sharing

[Nickolay V. Shmyrev]

There is additional penalty for distributed processing which makes it slower for such small datasets. For bigger sets you will see a speedup.

[rezaee]

what should be the numbers for 2 and 4 core machins?
CFG_NPART = ?
DEC_CFG_NPART=?

[Nickolay V. Shmyrev]

2 and 4.