However, how many hours does it run consecutively?

The start up is a trivial amount of total runtime.

I experience 4x to 8x slower perf running Java servlets on a brand new 1.4GHz T2 than on a 3 years old dual core AMD 2.2GHz both running Solaris 10. They will tell you that T2 is meant to do a lot in parallel but my loadtests is that you shouldn’t give it more than 100 simulaneous requests.

Also, I’ve experienced a lot more instability on JVM-Tomcat platform than on any Apache-PHP configs.
The problem here is that one bad webapp can pull all the JVM down, whereas in PHP, each request has its own http process (and memory space) which can die without hurting its neighbours.

You indicate that Guido is more or less saying “don’t use threads” and that he’s going in the wrong direction. I actually think he’s going in the right direction–

Threads make your program more complex and non-deterministic (and thus buggy.) I think the right direction is to use message passing concurrency.

Erlang is an interesting language that takes this idea to a high level:

http://www.pragmaticprogrammer.com/articles/erlang.html

Another interesting article on the problem with threads:
http://www.eecs.berkeley.edu/Pubs/TechRpts/2006/EECS-2006-1.pdf

Instead of (shared-state) multi-threaded programming which, can make your program more confusing and non-deterministic, both Erlang and Google Gears take the approach of using very efficient and light message passing for communication between basically isolated processes.

I think it’s the future and the right approach. Afterall, it’s how the Internet works… and the brain — both very complex systems that seem to scale a lot better than the “traditional” shared-state multi-threaded software development model.

I think Google is actually on the right track and obviously at the forefront of exploiting massively parallel system taking advantage of things such as map reduce (http://labs.google.com/papers/mapreduce.html.)

I’m an old time Java guy and I agree that Java is king, but I think it’s becoming more of a “systems language” with simpler scripting languages such as Python and JavaScript now taking on the brunt of applications development (because it’s EASIER!)

Leave the complex shared state multi-threaded program in the systems domain, and let the masses code in simple scripting languages using message passing concurrency so their programs are easy to understand and deterministic.

Glad to see Sun moving in that direction with the new scripting support and all the languages you mention that can run on the JVM.

I guess the bottom line is there is room for both points of view as one can be built on top of the other.

In the last few weeks I have been looking at different solutions and parallel python is the most impressive. Its a framework written in about 1000 lines of python code and it lets you easily scale “submit-job, retrieve-result” kind of applications from NUMA systems to a distributed memory clusters. Context switches are a concern, but I will hold judgment until I get around to benchmarking it properly.

http://www.parallelpython.com/