Wow, you have seriously no idea about High Performance Computing do you?!
First off, even the distance between buildings increases the latency of the network beyond what is useful for multi-node programs. Yes our campus has a wonderful fiber plant. No, it can't be used for HPC. Speed of light issues dude.
Next, keeping batteries under a cluster at full load is not very doable. Each rack pulls 13+kW, that's a ton o' battery. (Not to mention you also *need* to keep the cooling going during this time as well. Which on a good day is equal to the power consumed by the heat load.) Doable with standard loads on infrastructure type servers, not HPC nodes.
30'x30' room?! Hmm, no. Airflow around stuff is nice since that is what we use to cool the machines. Perhaps not the best plan, but plumbing racks for chilled water isn't doable at the moment.
And, where in the world do you get your $5k/node price??? You need to think again about what a cluster is and what such a beast is used for. Then, come back with a real number for us.
The majority of the cluster will use Ethernet for the networking. (All machines will connect to the *same* switch.) A small number will use our existing Infiniband infrastructure.
The machines will all have a single 160GB SATA disk, formatted with ext3, and run RHEL4. Half of the cluster will have 16GB of RAM and the other 32GB of RAM.
All clusters are served networked storage over NFSv3 from four BlueArc NAS devices.
Slashdot Mirror
Wow, you have seriously no idea about High Performance Computing do you?! First off, even the distance between buildings increases the latency of the network beyond what is useful for multi-node programs. Yes our campus has a wonderful fiber plant. No, it can't be used for HPC. Speed of light issues dude. Next, keeping batteries under a cluster at full load is not very doable. Each rack pulls 13+kW, that's a ton o' battery. (Not to mention you also *need* to keep the cooling going during this time as well. Which on a good day is equal to the power consumed by the heat load.) Doable with standard loads on infrastructure type servers, not HPC nodes. 30'x30' room?! Hmm, no. Airflow around stuff is nice since that is what we use to cool the machines. Perhaps not the best plan, but plumbing racks for chilled water isn't doable at the moment. And, where in the world do you get your $5k/node price??? You need to think again about what a cluster is and what such a beast is used for. Then, come back with a real number for us.
The majority of the cluster will use Ethernet for the networking. (All machines will connect to the *same* switch.) A small number will use our existing Infiniband infrastructure. The machines will all have a single 160GB SATA disk, formatted with ext3, and run RHEL4. Half of the cluster will have 16GB of RAM and the other 32GB of RAM. All clusters are served networked storage over NFSv3 from four BlueArc NAS devices.