[Beowulf] Parallel memory

Vincent Diepeveen diep at xs4all.nl
Tue Oct 18 15:02:58 PDT 2005


Hello Todd,

After rereading your posting. Do you know there is software like openssi?

http://www.openssi.org/cgi-bin/view?page=openssi.html

It might do exactly what you want to. It creates out of a cluster a single
shared system image which you can command just like as if it is 1 big
supercomputer.

Well 1 big partition from the 80s that is.

Assuming you neatly allocate your huge RAM in shared memory, each process
allocating a part of the memory and then cris cross attaching that shared
memory, if you do a memory access like that over openssi it will forward
in a very slow manner like 2048 bytes. 

The 4 bytes you're looking for you'll find then there.

This is ok if you are streaming in some sort of way.

Anyway you'll need to dramatically rewrite the application for it.

Keep in mind such systems like openmosix and openssi are beginner projects,
getting executed by big enthusiastic amateurs, who still own many machines 
with only 10 mbit cards and therefore you really should worry about those
1-2 milliseconds it takes to get you that say 2k bytes of memory.

If that's fine speed with you, take a look to openmosix and openssi,
and cash in your contract job money soon as soon as you got either of
those 2 systems to work.

Please take into account that migrating shared memory is not gonna happen
in a clever way with openmosix nor openssi like it is getting done on the
professional SSI boxes (though that also takes hours to migrate 1
applications shared memory correctly in IRIX, that is if you use a gigabyte
or 100, let's
not start discussing linux kernel extensions yet... for the altix box, and 
let's not start discussing the lightyears it would take on
openmosix/openssi with 10mbit interconnects and hubs from the 80s).

Please consider how many reads you are doing a second to remote memory.
Also consider that with many nodes odds are simply 99% you are accessing
remote memory in case you are accessing big arrays.

>From that you can calculate back what system time you at maximum want to
waste on a single access and how many bytes you need at once.

Vincent

At 01:09 PM 10/18/2005 -0500, you wrote: 
>>>>

font-family:Arial'>Are there any drivers, tools, etc. that can make the
memory space on a cluster look shared, something like pvfs but for memory?
I'm sure there would be a speed hit, but in this instance, speed isn't the
problem as much as memory.  We have a code that uses a huge amount of
memory and the memory usage is proportional to the cube of the problem
size, but the time for it to run isn't too much of an issue.  I've been
asked to parallelize the code using mpi which is going to be a major
effort.  However, I thought that if there was anyway, even if inefficient
speed wise, to create a virtual parallel memory system it would be better
than it using swap space and save a bunch of coding time. 

font-family:Arial'>  

font-family:Arial'>Also, are there any tools to help implement mpi in an
older code? 

font-family:Arial'>  

font-family:Arial'>Any thoughts? 

font-family:Arial'>  

font-family:Arial'>Thanks, 

font-family:Arial'>Todd 

font-family:Arial'>  _______________________________________________ 
Beowulf mailing list, Beowulf at beowulf.org 
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf 







More information about the Beowulf mailing list