Are you planning to use map/reduce within dotCloud? I ask because, otherwise, there'd be no reason to use HDFS instead of something with real filesystem semantics, comparable performance and better availability. Once such option is mentioned in the OP.
Disclaimer: I'm on the GlusterFS architecture team and advisory board, so I'm totally biased.
I agree Gluster seems like a more reasonable choice if you're starting from scratch. But if you have an existing hadoop cluster and want your app to interact with it, a simple way to mount HDFS would be great.