Where to begin? I am planning for a project to build a (very elaborate) all-round general purpose HPC cluster and I need some advice. This will be a complicated set up due to the fact that it involves two clusters being controlled from one master server. I plan to acquire a number of systems for this project. The following is a list of the systems I'm purchasing, their specs, and so forth:
Part I
-IBM BladeCenter E (Quantity = 6) http://www.redbooks.ibm.com/Redbooks.nsf/RedbookAbstracts/tips0995.html#contents
Each BladeCenter will have:
-14 IBM HS21 blade servers http://www-03.ibm.com/systems/xbc/cog/Withdrawn/bc_hs21_8853/bc_hs21_8853aag.html with the following specs:
-HP dx5150 (Quantity = 1) http://h18000.www1.hp.com/products/quickspecs/12105_na/12105_na.PDF
Since I happen to have one of these sitting around, and the server for the Kerrighed cluster doesn't have to be anything special, I decided to use this machine as the server for the aforementioned Kerrighed cluster, providing the NFS service for the node OS and all of the software. All of the blade servers will be connected via gigabit ethernet switch to this machine as well as the master server. This server will be running Ubuntu.
Part II
-HP Workstation Z800 (Quantity = 6) http://h18000.www1.hp.com/products/quickspecs/13278_na/13278_na.PDF
Each of these six workstations will have one or two NVIDIA Tesla C2075 GPGPU's. This will provide the supercomputer with some GPU processing power. These machines will be part of an MPICH2 cluster, and will run either Ubuntu or FreeBSD (I haven't decided yet).
-HP dx5150 (Quantity = 1)
I'm going to get another one of these to be the server for the MPICH cluster, and all of the Z800's will be connected to this server via gigabit ethernet switch. This will function as the head node for Part II.
Part III
-HP Proliant Microserver Gen8 (Quantity = 1) http://h18000.www1.hp.com/products/quickspecs/14565_na/14565_na.pdf
At the top of it all, there will be this machine. This will be the master server for both Parts I and II. Both clusters will be connected with yet another gigabit ethernet switch to this server, which will in turn, be connected to the internet. This server will be running FreeBSD, and is intended to provide the necessary services for all members of the cluster.
The whole supercomputer will be simply be these two clusters controlled via SSH from the HP Proliant MS G8. Using it should, in theory, be relatively simple once I get passwordless SSH set up for everything. I'm just not sure about a few things:
-Should I perhaps use something more powerful than a dx5150 as my servers for each nested cluster, or does it not matter for the servers?
-Same question as above, but for the HP Proliant Gen8?
-Does anyone have any recommendations as to how the Kerrighed cluster (Part I) could be controlled from the master server? I was considering using a tool like ClusterSSH or something, but I'm not sure exactly how to go about this.
I am all ears to suggestions about this.
Oh, and before you ask me "What in the world would you possibly need this monster machine for?":
3 Reasons:
1. I want to start a business that will provide HPC solutions for companies in all kinds of industries.
2. It will be a good learning experience for me.
3. I have way too much free time on my hands.
Edit: Oh, I just read up on Kerrighed, and it does have command line utilities for controlling the cluster which I can use to control it from the Proliant.
Part I
-IBM BladeCenter E (Quantity = 6) http://www.redbooks.ibm.com/Redbooks.nsf/RedbookAbstracts/tips0995.html#contents
Each BladeCenter will have:
-14 IBM HS21 blade servers http://www-03.ibm.com/systems/xbc/cog/Withdrawn/bc_hs21_8853/bc_hs21_8853aag.html with the following specs:
- Processors: Dual 2.33 GHz Quad-Core Xeon
- Memory: 4-8GB
-HP dx5150 (Quantity = 1) http://h18000.www1.hp.com/products/quickspecs/12105_na/12105_na.PDF
Since I happen to have one of these sitting around, and the server for the Kerrighed cluster doesn't have to be anything special, I decided to use this machine as the server for the aforementioned Kerrighed cluster, providing the NFS service for the node OS and all of the software. All of the blade servers will be connected via gigabit ethernet switch to this machine as well as the master server. This server will be running Ubuntu.
Part II
-HP Workstation Z800 (Quantity = 6) http://h18000.www1.hp.com/products/quickspecs/13278_na/13278_na.PDF
Each of these six workstations will have one or two NVIDIA Tesla C2075 GPGPU's. This will provide the supercomputer with some GPU processing power. These machines will be part of an MPICH2 cluster, and will run either Ubuntu or FreeBSD (I haven't decided yet).
-HP dx5150 (Quantity = 1)
I'm going to get another one of these to be the server for the MPICH cluster, and all of the Z800's will be connected to this server via gigabit ethernet switch. This will function as the head node for Part II.
Part III
-HP Proliant Microserver Gen8 (Quantity = 1) http://h18000.www1.hp.com/products/quickspecs/14565_na/14565_na.pdf
At the top of it all, there will be this machine. This will be the master server for both Parts I and II. Both clusters will be connected with yet another gigabit ethernet switch to this server, which will in turn, be connected to the internet. This server will be running FreeBSD, and is intended to provide the necessary services for all members of the cluster.
The whole supercomputer will be simply be these two clusters controlled via SSH from the HP Proliant MS G8. Using it should, in theory, be relatively simple once I get passwordless SSH set up for everything. I'm just not sure about a few things:
-Should I perhaps use something more powerful than a dx5150 as my servers for each nested cluster, or does it not matter for the servers?
-Same question as above, but for the HP Proliant Gen8?
-Does anyone have any recommendations as to how the Kerrighed cluster (Part I) could be controlled from the master server? I was considering using a tool like ClusterSSH or something, but I'm not sure exactly how to go about this.
I am all ears to suggestions about this.
Oh, and before you ask me "What in the world would you possibly need this monster machine for?":
3 Reasons:
1. I want to start a business that will provide HPC solutions for companies in all kinds of industries.
2. It will be a good learning experience for me.
3. I have way too much free time on my hands.
Edit: Oh, I just read up on Kerrighed, and it does have command line utilities for controlling the cluster which I can use to control it from the Proliant.