Wednesday, February 3, 2016

Cluster Building Fantasies

Me and the Clusters

When I was trying to put a bunch of  not-so-old computers  as my 1st cluster, I was a true novice who pushed himself into the beautiful and unknown world of UNIX/Linux. That was 11 years ago!. Nothing changed much: neither my naivety nor the no-so-old bunch of machines that I am stitching together to build my own cloud.

Basic Idea (Definitely NOT New!)

Create an environment on demand and on the fly that is customized for a specific task, as required by a user.
Examples of people who would need such an environment:
1)A physicist (like me) wants to run a simulation that requires a fairly powerful system for a certain period of time. In order to build the system (s)he would need to spend about $10K on the hardware alone and would need the services of a system administrator who would maintain the system ( students make good sys admins???) . Instead (s)he approaches someone like me, who has the means to setup that very specific environment that is needed to run that simulation.
2) A data mining firm wants to run a program for its data analysis and requires fairly powerful systems  and does not want to build a $50K system like mine, but rather use my system for a certain period of time fr a fraction of the installation costs.

A myriad other scenarios can be thought of.

Let us see if this is possible in theory and if it can be realized in practice and how well can this be realized, from a user perspective and of course from a business perspective


The logical parts of an on demand computational system

1.A system with many cpus stitched together to form a "supercomputer".
2. A large storage system ( right now I have about 120 TB of storage)
3. A database of different Operating Systems
4. A sufficiently fast network system
5. An authentication server to manage users

The list is definitely not exhaustive, but we will add more to the list in time, to make it more comprehensive.



Step 0 :  Building The Server from Parts from EBay

Budget < $10K as of Nov 2015

How do we stitch together a bunch of systems to form a massive computational giant like the ones at Amazon? While a lot of literature is available all over the internet, my attempt is to put together a document which looks at every little aspect of it, in one place.
Let us look at the available processing powers which are affordable. By affordable, I mean I do not want to spend more than the average processing costs (storage:$0.05/gb power: $15/cpu/year, etc.).

Servers of the Old!

When I was shopping for high-end servers, I chanced upon bunch of Dell Poweredge R900's. These were mostly decommissioned and were made available for under $1000.00 on eBay.
These are cheap (short term investment ) for someone who wants to learn a good deal of Hardware/ OS/Clustering/Virtualization etc.   Here is a link as a reference  :  Dell r900 on EBay .
These have come down quite a bit. One may also look at the early  R910's that are in the process of being decommissioned (!?!)
Armed with  four of these powerful power hungry servers, I went on to build my 1st personal cluster. Total of 16 Hexa-Core Xeon 7460  Processors, 512 GB PC2-6400 DDR2 ( 800MHz) Memory. Fortunately all of them came with a Dell perc 6i which is just the PowerEdgeRAID Card that can handle various types of RAID configuration RAID0, RAID1 RAID5, RAID 10 etc.
Continuing my hunt on EBay for the largest possible capacity drives that i could afford and that would recognized by the RAID card, I found a lot of 300GB 10K SAS for  $40.00 a piece. Gold mine!
I ordered 16 of them: 4 per machine.
Here are the specs of my hardware

 Dell PowerEdge R900 
______________________________
4x Xeon 2.66GHz Six Core X7460
128 GB PC2-6400 DDR2 SDRAM
4x300 10K RPM 3.5" Internal HDD
2x4 port Gigabit Ethernet Adapter 
______________________________

I put them together in a makeshift  rack (Book Shelf) till I could save enough money to get a decent rack.

Now three very serious annoyances here! first,the noise is more than that a jet plane taking off and the other one is each of them draws a peak power of of 1200W. But most importantly, these 16 processors generate a lot of heat!  So much that I did not need a heater for the entire house for most of the  Jersey winter.

In the next post I will write about the OS installation, network and the clustering process.