[Beowulf] Rackable / SGI

[Greg Lindahl]

On Fri, Apr 03, 2009 at 08:15:06PM -0400, Joe Landman wrote:

> It has been, realistically, the only serious choice for big/fast  
> data systems on Linux for quite some time.

Most big/fast data systems don't have big filesystems. In HPC, the
biggest use something like Lustre; in business, the biggest use
something like BigTable/GFS. In both cases the underlying filesystems
don't have to be that big.

Now for what you really meant ("big filesystems"), that was pretty
true when XFS first came out. But Emacs (oops, I meant ext2) was
improved upon subsequently, and the remaining differences are more a
lifestyle choice than a clear technical choice.

> Ext3 has some serious  
> performance limits due to its journaling design.  Never mind its other  
> issues.  There was mention of this in this past week's LWN.net.

If you use the default desktop config, yes. This is a common situation
in lifestyle arguments: "performance would be poor if you use product
X like an idiot." It's an ease-of-use problem.

> I wouldn't use ext3 for anything other than small partitions (100 GB or  
> so).  Too many cases seeing the fsck need to get triggered for some  
> reason ... the wait is horrible.

I use crash-only techniques on my cluster, and I've never seen a
spurious fsck. I agree that it will be nice that ext4 makes fsck times
much faster, but that isn't bothering me.

> Zfs is not the revealed word of some deity, in file systems.  This mind  
> set is painful to deal with, and often winds up with people having  
> *very* unrealistic expectations of what it is, what it can do, and how  
> it performs.

Yeah, it almost reminds me of when XFS came out. (Did I say that out
loud?)

You're probably a vi user :-)

-- greg