Commodity components for Cloud

2 views
Skip to first unread message

Khazret Sapenov

unread,
Aug 26, 2008, 11:44:54 PM8/26/08
to cloud-c...@googlegroups.com
Interesting concept of "... creating a datacenter that will have scalable interconnection bandwidth, making it possible for an arbitrary host in the datacenter to communicate with any other host in the network at the full bandwidth of its local network interface."
 
They also claim, that "From a cost perspective, to build out a 25,000 node cluster today using current techniques with 100 percent bandwidth, just the switching equipment would cost somewhere in the order of $28 million, whereas with our technique using the identical network elements, would deliver the same performance but incur costs of maybe $4 million. That's a factor-of-seven difference."
 
original article:
 
Looks like a trend (of using commodity components for large scale deployments), started in Google and confirmed several times by other companies. It can be explained by such factors as capex/opex reduction and general human conservatism (opposition to change and innovation in terms of using non-standard components). 

Bob Lozano

unread,
Aug 27, 2008, 12:12:12 AM8/27/08
to cloud-c...@googlegroups.com
There is no doubt that datacenter components will evolve / devolve to commodity components - the economics are simply too compelling at this point, at least for the foreseeable future.

Probably the hardest architectural component to consider commoditizing is the switching interconnect - it can be sorely tempting to not want to think of network topologies, if for no other reason than it can simply be a pain.

While the article is very short on details, you can download the paper itself here. A litle too much to thoroughly consider right now, but it does look interesting at first glance.

In any case, I think any improvements we can make in interconnects (particularly where the building blocks themselves remain commodity) is a great thing. Who could argue any other way?

Having said that, I think for many apps (outside of the non-chunkable mega-simulations/models/etc - you know who you are) even bigger gains are to be had from having the cloud app layer 1) know how to keep work moving to the data, and 2) self-organize computing and storage elements around things like the commodity switch boundaries.

That way you can let the "cloud enablement" teir organize the work appropriately, without burdening the app developer.

Bob
--
Bob Lozano

Appistry
www.appistry.com/blogs/bob (professional blog)
hopeitis.com (personal blog)
Reply all
Reply to author
Forward
0 new messages