Jonathan Seidman 2009-10-21, 16:15
Dhruba Borthakur 2009-10-21, 18:22
gmackey@... 2009-10-22, 16:37
On Oct 22, 2009, at 9:37 AM, <[EMAIL PROTECTED]> wrote:
> As with Dhruba's comment, so long as it is just the namenode that is
> running on a networked file system everything should be chill. The
> keeps all of its working metadata in main mem, and it only
> pushes a log file out to hard storage (and if I remember correctly
> you can
> adjust this time window in one of the site files).
Actually it pushes out the update logs on each and every update
The checkpoint however is pushed out periodically.
Also, at yahoo, we push out NN state to multiple disks and one of the
"disks" is a nfs filer. This is configurable.
> However, you are going to run into huge performance issues running
> datanodes over a networked storage system. Having to push that many
> requests over a network for a respectable mapreduce job is going to
> your equipment.
> - Grant
> On Oct 21 2009, Jonathan Seidman wrote:
> >Apologies if this has been answered previously, but I'm unable to
> >anything that seems to cover this.
> > It's clear that datanodes require local storage for Hadoop to
> > efficiently, but is there any significant disadvantage to using
> > storage for namenodes? We're exploring the possibility of using a
> > different class of hardware for our namenodes with attached
> storage and
> > little or no internal storage. Some of the benefits this would
> provide us
> > are: 1) allowing our sysadmins to deploy hardware that they're
> > with and already have considerable experience keeping up in a
> > environment. 2) no namenode downtime to replace a failed disk.
> >We don't anticipate that this approach would cause any significant
> >degradation to performance, but let me know if there's something
> we're not
> Grant Mackey
> PhD student Computer Engineering
> University of Central Florida
> Rm 231 cube 5 (321) 960-8851
Grant Mackey 2009-10-22, 19:16