Boston Linux & Unix (BLU) Home | Calendar | Mail Lists | List Archives | Desktop SIG | Hardware Hacking SIG
Wiki | Flickr | PicasaWeb | Video | Maps & Directions | Installfests | Keysignings
Linux Cafe | Meeting Notes | Blog | Linux Links | Bling | About BLU

BLU Discuss list archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]


Hash: SHA1

markw at wrote:

| Jon Masters Wrote:

|>Best stick with one of the suggestions already made. NUMA is designed
|>for real physical hardware, what you want is a cheap cluster.

| Actually, I have a cheap cluster. I have five dual CPU SMP boxes networked
| on a high speed switch. I've done several projects that use this setup as
| either a load balanced server or a beowulf cluster.

Does that count? ;-)

| While I haven't used Mosix (I use these machines for tests and
| development, I hesitate to use a heavily modified kernel)

I think MOSIX is a cute hack. But that's all it is. In 2.6 it's even
easier to bodge process migration queues and shoved processes from one
node to another dyanamically, but the vm cost is so high that it's
counter-productive to migrate under many circumstances.

| The reason why I'm curious about NUMA simulation is the idea that
| it could be possible to run programs written similarly to how you'd write
| MPI programs, but with "implicit" sharing of memory instead of "explicit"
| sharing in the form of messages.

I'd like to see something like BSP revived and done properly to support
NUMA,SMP, and clusters built on regular networks - it'd take a bit of
kernel support and be slow, but a good academic exercise for someone
with some free time.

| ccNUMA should be able to be implemented on the x86 architecture using
| virtual memory mapping and the shm library. It would probably take a
| kernel module or two, but it would be cool.

It has likely been done already. Certainly we were talking about it on
kernelnewbies a while back - but I think it'd be so horribly inefficient
as to make the task little more than a semi-cool exercise in abstraction
- - many others have pointed out that for performace we don't always want
to make it easy for anyone to write seemingly "high performance" code.
That'd be akin to Javaising the whole process :-)

Version: GnuPG v1.2.4 (GNU/Linux)
Comment: Using GnuPG with Thunderbird -


BLU is a member of BostonUserGroups
BLU is a member of BostonUserGroups
We also thank MIT for the use of their facilities.

Valid HTML 4.01! Valid CSS!

Boston Linux & Unix /