Network Booting
I'm going to be building a linux cluster soon and I want to boot linux from a network server to cut down on the cost per node. Is there any certain distributions of linux that specialize in this, or is this ability included in all distributions. Would 1GB of memory on the server side, and 256MB on the client side be sufficent?
Right now the cluster is only going to be 3 nodes and a server, but the cluster may expand to as many as 100 nodes in the future.
I recall once trying a modified version of Knoppix ("ClusterKnoppix" IIRC) that did this but I'm sure there are others too. Search DistroWatch and Google a bit and you'll surely find some.
Quote: Original post by I_Smell_Tuna
I'm going to be building a linux cluster soon and I want to boot linux from a network server to cut down on the cost per node.
Almost pointless, although a lot of clusters do network boot, but not for that reason. The nodes are pretty expensive, but still cheap compared to the annual salary of a Linux cluster sysadmin.
Linux clusters usually network boot for management reasons, i.e. still have local HDs for swap/temp space and possibly storing system software (although that might conceivably be via NFS).
Quote:
Is there any certain distributions of linux that specialize in this, or is this ability included in all distributions. Would 1GB of memory on the server side, and 256MB on the client side be sufficent?
Most distros will work in a network boot configuration. Debian is popular.
On the clients, you will need as much memory as your application requires.
Most scientific applications require absolute GOBS of memory.
I assume you have factored in your applications' memory usage into this. 256Mb is hardly any. In fact, I wouldn't recommend less than 2G on the clients. The server won't really need much memory I guess.
Quote:
Right now the cluster is only going to be 3 nodes and a server, but the cluster may expand to as many as 100 nodes in the future.
Well, do your sums. Fewer nodes with better performance is always better than more nodes with worse performance.
Ask your users how well their applications scale. Benchmark them on nonclustered hardware or small clusters for testing. Your software developers may not be capable of exploiting multiprocess systems very well. Bear this in mind.
Comms is also a big factor. If you have a zillion Ghz AMD64 with 100G of RAM, it's no good if it spends 90% of its time waiting for the network.
You could find that your money would be better spent on fast networking than lots of fast nodes. Don't waste your precious budget!
Mark
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement