Windows HPC Server 2008 R2 Hits Second Beta

Status
Not open for further replies.

ivan_chess

Distinguished
Mar 11, 2010
61
0
18,630
Sounds interesting but HPC OSs need to be really small. Compute nodes don't have hard drives so the OS must be sent over the network and stored entirely in the RAM. Feature filled Windows won't be small.
 

p05esto

Distinguished
Jun 11, 2001
876
1
18,980
You have to give it to MS, they are really turnign up the heat the past few years. They were a little stale and comfortable, someone woke the sleeping giant.
 

rbarone69

Distinguished
Aug 16, 2006
241
0
18,690
[citation][nom]ivan_chess[/nom]Sounds interesting but HPC OSs need to be really small.[/citation]

To start I have to disclaim that I currently harness one of man's greatest known reserves of HPC ignorance...

MS has done a pretty good job on the Server 2008 series to strip it down to a very small featureless footprint. I'd bet they have taken this a step further for this particular application.
 

ivan_chess

Distinguished
Mar 11, 2010
61
0
18,630
[citation][nom]rbarone69[/nom]To start I have to disclaim that I currently harness one of man's greatest known reserves of HPC ignorance...MS has done a pretty good job on the Server 2008 series to strip it down to a very small featureless footprint. I'd bet they have taken this a step further for this particular application.[/citation]

There are some resources out there, here are some undergrad projects that cover the basics of HPC:

http://institutes.lanl.gov/isti/summer-school/cluster-student-projects/
 
Well depending how their running this thing, it might not be a server "OS" in the typical notion. Now I don't work on this, but typically HPC's have "controller" node's and slave nodes with the controller nodes having some form of storage and feeding the slave nodes their boot code / data / jobs. You could do the same thing here, load the MS Server OS on a controller node, then remote boot all the data nodes from the controller node, feeding them only a minimal OS. Try not to think of each node having a full MS Windows with gui / driver infrastructure, only a small kernel.

I'm liking the idea about integrating with Windows 7 to use spare client CPU cycles. When you start to look at typical offices, imagine 100 client's giving space cycles during the night or even during the regular day.
 

MrKKBB

Distinguished
Jul 7, 2008
44
0
18,530
[citation][nom]ivan_chess[/nom]Sounds interesting but HPC OSs need to be really small. Compute nodes don't have hard drives so the OS must be sent over the network and stored entirely in the RAM. Feature filled Windows won't be small.[/citation]

Really? Most of the clusters I have worked with do have their own local drive. The one thing I always wondered about is the price. A $200/node OS is a lot more than the free Rocks using CentOS, especially when you have a 1000 nodes!
 

ivan_chess

Distinguished
Mar 11, 2010
61
0
18,630
[citation][nom]MrKKBB[/nom]Really? Most of the clusters I have worked with do have their own local drive. The one thing I always wondered about is the price. A $200/node OS is a lot more than the free Rocks using CentOS, especially when you have a 1000 nodes![/citation]

Large clusters would have a hard drive failing every 5 minutes. Check out my link to the LANL summer institutes above for more details on diskless nodes and management. We used CentOS too and you can push that out to nodes without disks (after removing unnecessary features). Storage arrays are more expensive but they are more failure tolerant than the cluster. You lose money when your cluster goes down so the storage array is cheaper in the long run.
 

randomizer

Champion
Moderator

It would be like asking if Linux is going to do battle with Windows in the home desktop market. Both OSs dominate their respective markets by such a large margin that they are unlikely to be shaken any time soon. And let's not forget those licencing fees with Windows HPC that Linux doesn't have (ignoring commercial support costs if the company went with that option).

From the article:

Interoperability options. We’ve heard from customers that “rip and replace” isn’t often a viable option for building out their clusters.

I hope they didn't spend much money researching this blatantly obvious fact.
 
Status
Not open for further replies.