Comments on: You Are Not Supposed To Do That http://www.linux-mag.com/id/7220/ Open Source, Open Standards Sat, 05 Oct 2013 13:48:18 +0000 hourly 1 http://wordpress.org/?v=3.1 By: shochschild http://www.linux-mag.com/id/7220/#comment-6040 shochschild Wed, 30 Nov -0001 00:00:00 +0000 http://www.linux-mag.com/id/7220/#comment-6040 In order for HPC to fully take advantage of commodity products, some cherished HPC traditions might need to be reconsidered. For example, not only Fortran, but MPI and OpenMP as well.<br /> <br /> I work for a commercial vendor of a developer product specifically focused on high performance data intensive analytic application development. <br /> <br /> We exhibited at SC'08 and had fun, but essentially, we were met with disbelief when the attendees realized that we were proposing that they use Java on 32-core SMP boxes. Java? SMP? Heresy!<br /> <br /> steveh<br /> www.pervasivedatarush.com In order for HPC to fully take advantage of commodity products, some cherished HPC traditions might need to be reconsidered. For example, not only Fortran, but MPI and OpenMP as well.

I work for a commercial vendor of a developer product specifically focused on high performance data intensive analytic application development.

We exhibited at SC’08 and had fun, but essentially, we were met with disbelief when the attendees realized that we were proposing that they use Java on 32-core SMP boxes. Java? SMP? Heresy!

steveh
http://www.pervasivedatarush.com

]]> By: coutinho http://www.linux-mag.com/id/7220/#comment-6041 coutinho Wed, 30 Nov -0001 00:00:00 +0000 http://www.linux-mag.com/id/7220/#comment-6041 As said in http://www.linux-mag.com/id/2292, MPI and OpenMP can have their limitations and should be replaced by something better.<br /> <br /> Maybe I am a being conservative here, but Java has some limitations to be accepted in HPC:<br /> <br /> There a lot of disbelief due to be a interpreted language. I know that most of the time JITs can achieve good performance, but they can't do heavy optimizations. Perhaps if someone make a vectorizing JIT, people will take Java more seriously.<br /> <br /> Sometimes explicit memory management can achieve huge impact on performance.<br /> In this paper they show that a GC program runs as fast as explicit managed memory if the machine had three times more memory [http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.61.9682]. Most people wold think "ok, just buy more memory", but HPC people think "Given a cluster with X GB of memory, what is the speedup if I use explicit managed memory?" As said in http://www.linux-mag.com/id/2292, MPI and OpenMP can have their limitations and should be replaced by something better.

Maybe I am a being conservative here, but Java has some limitations to be accepted in HPC:

There a lot of disbelief due to be a interpreted language. I know that most of the time JITs can achieve good performance, but they can’t do heavy optimizations. Perhaps if someone make a vectorizing JIT, people will take Java more seriously.

Sometimes explicit memory management can achieve huge impact on performance.
In this paper they show that a GC program runs as fast as explicit managed memory if the machine had three times more memory [http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.61.9682]. Most people wold think “ok, just buy more memory”, but HPC people think “Given a cluster with X GB of memory, what is the speedup if I use explicit managed memory?”

]]>