[thomasjames]

Does anyone know what the effective bandwidth and latency would be between machines? They talk about thousands of cores a little too casually. I guess they mean unrelated independent instances of some task as opposed to actual parallelism. I am confused as to what market segment they are going for here.

[ldargin]

According to this review, their bandwidth is high and latency low, compared to EC2: http://bit.ly/Z3Peob

[andrewem]

Non-shortened link: http://gigaom.com/2013/03/15/by-the-numbers-how-google-compu...

[thomasjames]

That is pretty fast. Obviously not like infiniband or what you use for message passing, but pretty considerable. It would be difficult to scale to thousands of nodes that need to communicate, but it is definitely good for most applications. When they mentioned the biostatistics group, I thought they might have been making a pass at scientific HPC type applications, but I guess this is more specialized. Prices are really reasonable, too.

[jgreen10]

http://blog.zencoder.com/2012/07/23/first-look-at-google-com...

[ldargin]

It looks like Google improved things a bit in the 7 months since that review.