Azul Means (Big) Blue: There's a new kind of mainframe coming and it isn't from IBM

[Readmore]

http://www.pbs.org/cringely/pulpit/2008/pulpit_20080229_004404.html

[mixmax]

I'm amazed that IBM isn't in trouble - their services and products seem so expensive that I wonder how anyone could ever buy them. I just came home from a dinner with somebody working for IBM, and she told me that getting a new customer into their CRM costs around $5000 and takes a week!

A month back a story ran in Denmark where IBM had bid to install an Oracle database, and wanted $30.000 for doing so. A small company made fools out of them by having an engineer do the same thing in less than half an hour. In a straitjacket, only using his nose. Video here: http://www.youtube.com/watch?v=CHzV4LZnvHc

[xirium]

768 custom processors with hardware garbage collection in a 14U case is quite impressive. I'd given up all hope of seeing a Java mainframe like this 10 years ago.

Regardless, I'd like to see this benchmarked against a PlayStation3 cluster ( http://news.ycombinator.com/item?id=125829 ). If you can cope with the 512MB memory then I suspect that the PS3 offers better value for money.

[wmf → xirium]

Why benchmark apples against oranges? Any app that will run well on a PS3 will suck on Azul and vice versa.

[wmf]

Some thoughts on Azul:

The comparison of Azul to a mainframe is apt, since there is no low-end Azul box. If you don't have $50K to spend, go home. Startups can stop reading now.

The network-attached coprocessor idea doesn't give me a good feeling; it sounds like yet another point of failure.

Because Azul uses hundreds of cores, most apps won't scale out of the box; Azul-specific tuning like lock splitting is needed.

I want to like Azul but in the end it looks like a niche box.

[ggruschow → wmf]

There's more to Azul than just the fact they have a zillion cores in a box.

They've also done some awesome work on optimistic thread concurrency and concurrent garbage collection.

[aswanson]

What about native OS SMP support? Does this architecture easily map and scale linux or freebsd? Seems like the dedicated engines for java should be replaced with TCP offload engines and let the server developers code in whatever language they choose. I don't like it, too tied to a language choice.

[wmf → aswanson]

I can understand that viewpoint, but porting the whole Linux universe to a new architecture is a much larger job than porting a Java VM.