11:07 AM
me: hi. is there some progress on kill task issue?
11:08 AM
fdmanana: hi

no :(

only read and understood most of hadoop's jvm task umbilical interface/protocol
11:09 AM
me: I see. i'm going to re-implement the 'BSPTaskRunner' class.

fdmanana: why?
11:10 AM
as far as I remember it's a small class, no?
11:13 AM
me: in hadoop, tasks are executed on seperate JVMs from TaskTracker,
with some java arguments.
11:14 AM
fdmanana: right

including the task log dir, task id, jvm id, etc

me: currently, BSPtask runs on GroomServer JVM.

yup.

it also seems related with kill method.

right?
11:15 AM
fdmanana: right

me: OKOK.

fdmanana: it's easier to kill abruptly a jvm (child os process) rather
than a java thread

me: yes..

fdmanana: once the child jvm is spawned, it knows what to do, since we
pass it the job class/jar name etc
11:16 AM
also the child jvm can taslk to the groom server through rpc

(to report progress, and report completion for e.g.)
11:20 AM
me: yes. today, i can work on BSPTaskRunner. i think, it's better to
separate it to small tasks. is it OK?

fdmanana: yes, ok for me :)

me: thanks!

;-)

-- 
Best Regards, Edward J. Yoon
http://blog.udanax.org
http://twitter.com/eddieyoon

Reply via email to