[Starcluster] some questions about starcluster ganglia, home and ebs

[Justin Riley]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Hi Markos,

Sorry for not responding sooner, this one got lost in my Inbox. I've
copied the list on this one, hope you don't mind.

> Do you remember I told you that we want to use ganglia? Well, we decide
> to use, we installed and it works!

Awesome, glad to hear that. Was there anything special involved besides
apt-get installing it?

>but we have a problem...
> problem:
>  - We launch an AMI with ganglia and starcluster
>  - We run starcluster start with 2 nodes
>  - these 2 nodes apparently we can't access directly for example via
> telnet or a web browser, I mean the only way is through the starcluster
> commands for example "starcluster sshmaster mycluster"...

This is likely related to the fact that you need to open up port 80 in
the cluster's security group. StarCluster creates a new security group
for each cluster it starts using the cluster's tag name prefixed with
'@sc-'. You can easily change the security group permissions in the AWS
management console or ElasticFox. Thanks for bringing this issue up,
I'll consider adding support to the config for specifying ports to open
after a cluster has been created. For now manually configuring security
group permissions yourself will have to do.

> we realized that the starcluster AMI has 10GB
> of partition for /home even if we launch a medium or xlarge instance,
> and for our output, we require more than 20GB, so well, we mount an EBS
> and it was solved... But we don't know if it's possible that the /home
> directory can be bigger so in that way we don't have to use EBS? I saw
> that /mnt directory has a lot of space, can we repartition that?

So, first of all is this just loads of temporary storage that you don't
plan on keeping around? If you plan to keep it, EBS is a must. If not,
you could certainly reconfigure the space mounted on /mnt to /home. But
then I'd ask, why not just write data to /mnt either directly or by
creating a symlink in /home/user to /mnt? Moving the space on /mnt to
/home might be something I want to do. I'll have to think on that. The
usual model is to just use an EBS volume instead but

> And the last question Justin, Does the jobs that we run takes more
> time using EBS? I mean, is there a difference in time between using or
> not EBS?

That's a good question and unfortunately there's not really a cut and
dry answer. I've read some arguments both ways about whether local
storage or an EBS volume is faster. I found the post below to be the
most informative:

http://developer.amazonwebservices.com/connect/thread.jspa?messageID=125197&#125197

In this post an Amazon employee responded that:

"The ephemeral drives have a pretty significant first-write penalty
....
Finally, as I've written before, if you are mostly doing sequential I/O
(which is what dd measures), you're not going to see much difference
between an EBS volume and a single ephemeral store, and 4 ephemeral
stores striped together can give significantly higher throughput for
sequential writes than EBS."

The same employee commented that it also might be possible to improve
performance using striped EBS:

http://developer.amazonwebservices.com/connect/thread.jspa?messageID=132387&#132387

In any event, I would suggest using dd/bonnie to actually measure this
yourself and see what you're getting on average.

Hope that helps,

~Justin

On 05/16/2010 01:36 PM, markos calderon wrote:
> Hi Justin,
> 
> Do you remember I told you that we want to use ganglia? Well, we decide
> to use, we installed and it works! but we have a problem... This is the
> problem:
>  - We launch an AMI with ganglia and starcluster
>  - We run starcluster start with 2 nodes
>  - these 2 nodes apparently we can't access directly for example via
> telnet or a web browser, I mean the only way is through the starcluster
> commands for example "starcluster sshmaster mycluster"... 
> 
> We try to install even if a web server in this two nodes but we don't
> have success, I imagine that this requires a change in the code, I don't
> know if you can help us with some guidelines, so in that way we can	
> change it...
> 
> Another thing Justin, we are trying some examples with starcluster that
> the output is very big... we realized that the starcluster AMI has 10GB
> of partition for /home even if we launch a medium or xlarge instance,
> and for our output, we require more than 20GB, so well, we mount an EBS
> and it was solved... But we don't know if it's possible that the /home
> directory can be bigger so in that way we don't have to use EBS? I saw
> that /mnt directory has a lot of space, can we repartition that? 
> 
> And the last question Justin, Does the jobs that we run takes more
> time using EBS? I mean, is there a difference in time between using or
> not EBS?  
> 
> Thanks in advanced for your reply...
> 
> -- 
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> Markos Calderon
> Miembro de CTI y Taws
> Celular: 080908111
> Skype: markos.calderon
> Facebook: http://www.facebook.com/markos.calderon
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> 

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2.0.14 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/

iEYEARECAAYFAkv0ij4ACgkQ4llAkMfDcrlxRgCfT/IEBoTaG+vziYDhRLi9x1p5
FyUAnidO3B0w+gkW/34TSiVf520SHqIw
=hXMF
-----END PGP SIGNATURE-----