Not too worried. I could of course have only the two other machines running as 
both Master and Slave for instance, but if I could incorporate the desktop into 
the cluster as well I think that would be preferable so that the other two can 
run as complete slaves and I can offload all of my computation jobs to them 
through Mesos. Since this won't be backing any web applications, websites, or 
anything like that, I don't expect anything to be getting hammered. I was more 
thinking if there was anything particular about the Mesos Master process that 
might conflict when it is being run on essentially a non-dedicated server and 
not isolated (as is the case when setting up a local dev mesos cluster with 
Vagrant or the like). Of course I could always stick Mesos Master inside a VM, 
Vagrant, or Docker container on the workstation.

Just trying to get a handle on any potential pitfalls. All of the great 
examples out there for building a cluster are either cloud-based or a VM'd 
cluster on a single local machine and not much info when it comes to building a 
physical cluster.

From: Aaron Carey [mailto:[email protected]]
Sent: Tuesday, July 21, 2015 12:05 PM
To: [email protected]
Subject: RE: "Cluster of Workstations" type design for a Mesos cluster

There's nothing stopping you running the mesos master and slave process on the 
same machine, so you could run the master process on your non-desktop machine 
if you're worried.

We have the master and slave processes run as docker containers and they can 
both end up on the same machine without any problems.
________________________________
From: Gaston, Dan [[email protected]]
Sent: 21 July 2015 14:44
To: '[email protected]'
Subject: RE: "Cluster of Workstations" type design for a Mesos cluster
Is there likely to be any issues with the Master? Given it would be an active 
desktop it would be running all of the typical mesos master stuff, plus say an 
active Ubuntu desktop environment. It would also need to host things like a 
local Docker registry and the like as well, since the compute nodes wouldn't 
have direct access to the wider internet.

From: [email protected]<mailto:[email protected]> 
[mailto:[email protected]] On Behalf Of Jeff Schroeder
Sent: Tuesday, July 21, 2015 10:42 AM
To: [email protected]<mailto:[email protected]>
Subject: Re: "Cluster of Workstations" type design for a Mesos cluster

As far as mesos is concerned, compute is a commodity. This should work just 
fine. Put Aurora or Marathon ontop of mesos if you need a general purpose 
scheduler and you're good to go. The nice thing is that you can add additional 
slaves as you need. I believe heterogeneous clusters are best if possible, but 
absolutely not a requirement of any sort.

On Tuesday, July 21, 2015, Gaston, Dan 
<[email protected]<mailto:[email protected]>> wrote:
Let's say I had 2 high-performance workstations kicking around (dual 6-core, 
2.4GHz, xeon processors; 128 GB RAM each; etc) and a smaller workstation 
(single Xeon 4-core, 3.5GHz and 16 GB RAM) available and I wanted to cluster 
them together with Mesos. What is the best way of doing this? My thought was 
that the smaller workstation would be at my desk (the other two would be in the 
same office) because it would be used for development work and some general 
tasks but would also be the master node of the mesos cluster (note that HA 
isn't a requirement here). This workstation would have two NICs, one connected 
to our institutional network and the other making up the private network 
between the clusters.

Is this even doable? Normally you would have some sort of client submitting to 
the Master but in this case the Master node would be serving up multiple roles. 
The other workstations would probably not have access to the institutional 
network, so all software updates and the like would have to be piped through 
the master workstation. There would also be a relatively large NAS device 
connected into this network as well.

Thoughts and suggestions welcome, even if it is to tell me I'm crazy. I'm 
building a small scale compute "cluster" that is fairly limited by budget (and 
the needs aren't high either) and it may not be able to be located in a 
datacenter, hence the cluster of workstations type setup.



[NSHA_colour_logo.jpg]

Dan Gaston, PhD
Clinical Laboratory Bioinformatician
Department of Pathology and Laboratory Medicine
Division of Hematopathology
Rm 511, 5788 University Ave.
Halifax, NS B3H 1V8





--
Text by Jeff, typos by iPhone

Reply via email to