perfSONAR User Q&A and Other Discussion

[Mark Feit <>]

Eli Dart writes:

 

It would be interesting to have a conversation with Amazon about expected data transfer performance numbers for different types of AWS instances, using different technologies (HTTP, Globus S3 connector, etc).

 

Do you think they might be receptive to something like this?

 

I hate to be the thrower of cold water, but it’s likely to be a very short conversation.

 

Unlike R&E, which has every reason to share this kind of thing in the interest of furthering research; businesses have every reason to play their cards very close to their chests for competitive advantage.  They also tend to be intentionally vague about performance because the oversubscription-based business model makes it unpredictable.  (Measurement in a rent-a-box environment has a long list of pitfalls and will probably be the subject of something I write in the future because it seems to be of interest to a lot of people.)

 

There would be no business case for installing dedicated measurement machines because the cost relative to the return wouldn’t be worth it.  They may do this internally, but it’s done as part of their monitoring infrastructure and isn’t going to be available for us to run perfSONAR.   I should also note that Amazon does have a partnership with Rackspace to offer instances on dedicated systems, but they’re (predictably) expensive and probably don’t share EC2’s infrastructure.

 

One other thing to consider:  Some service providers have language in their user agreements that forbid disclosure of observed performance to third parties.  It would be wise to consult your agreements with Amazon and others before having these discussions in any detail or making rented instances part of a publicly-visible dashboard to avoid running afoul of the terms.

 

--Mark