Re: IBM Unleashes World's Most Powerful Server

____________________________________________________________________________
Evan Harris wrote:

Hi Charles

Somehow you have to beat the law of diminishing returns.

Each additional server in a cluster requires processing overhead just so it knows the state of the other machines in the cluster, and any work the cluster is performing. As the number of servers in the cluster increases the overhead of maintaining the state of the cluster on each individual node increases leaving less processing available for real work. If you have a database spread across 61 servers it's a lot of work (!) just to keep the database up to date every time a transaction occurs.

Of course in Google's case (as someone else mentioned) this is no doubt one of their prime business assets - the knowledge of how to do this. In their case though I'd hazard a guess that having many servers is as much driven by the need for many simultaneous connections and ensuring that there is no bottleneck at the network interface level as opposed to a requirement for raw processing power and storage capacity.

Naturally I'm speaking from a theoretical perspective rather than one of first hand knowledge and experience :)

Regards
Evan Harris



At 05:23 a.m. 26/07/2006, you wrote:

An additional interesting thought:

One of these 595:
http://www.tpc.org/tpcc/results/tpcc_result_detail.asp?id=106072501
System Information
Total System Cost       11,967,178 US $
TPC-C Throughput                4,016,222
Price/Performance       2.98 US $


Vs.

Sixty-one of these Dell Poweredge servers:
http://www.tpc.org/tpcc/results/tpcc_result_detail.asp?id=106063001
System Information
Total System Cost       64,512 US $
TPC-C Throughput                65,833
Price/Performance       .98 US $


In other words, what would the Price/Performance be of a cluster of the
lowest priced individual systems with the performance of the most
powerful single system.