Network Memory System proposition
Yuan-chi (Bill) Chiu
If you have a Ethernet connection or any fast network connection to
a UNIX system, the following may not be such a bad idea:
Run a server on your UNIX account that malloc() UNIX system memory
and allow a client running on AmiTCP side to use those memory over the
network!
If this sounds like parasite, you're right :-) But it could be
one possible alternative to virtual memory systems. Simply write a
server that process memory requests, malloc() a block of memory
from the UNIX system, and let the client program read/write this
piece of memory. A client program on AmiTCP side will do what most
virtual memory programs do except that it gets its memory resource
from the server across the network instead of a cache on disk.
Possible? Implementable? I think so :-)
--
+-------------------------------------------------------------------------+
| Yuan-chi (Bill) Chiu. hbcs...@csun.edu Amiga Enthusiast. |
|-------------------------------------------------------------------------|
| A process cannot be understood by stopping it. Understanding must move |
| with the flow of the process, must join it and flow with it. |
+-------------------------------------------------------------------------+
Michael D. Bayne
: Run a server on your UNIX account that malloc() UNIX system memory
: and allow a client running on AmiTCP side to use those memory over the
: network!
So how would you access this memory on the Amiga side? If I wanted to assign a
value to address 16, but address 16 was on the Unix machine, would you have an
MMU table set up such that a page fault was generated and this program would
grab that page of memory from the Unix machine and stick it somewhere in local
Amiga memory? Sounds like GigaMem with the swapfile on an NFS mounted
Unix-side RAM drive.
It oozes slowness.
--
/o)\ Michael D. Bayne
\(o/ mba...@qualcomm.com
Stephen D. Holland
>Yuan-chi (Bill) Chiu (hbcs...@huey.csun.edu) wrote:
>: Run a server on your UNIX account that malloc() UNIX system memory
>: and allow a client running on AmiTCP side to use those memory over the
>: network!
Distributed Shared Memory is actually a hot topic in OS's these days.
Its a neat communications paradigm, but still needs research work.
I think for what you want you should just get a bigger local drive and run
virtual memory.
-Steve
--
---------------------------------------------------------------------
Steve Holland N1RAO "I never let my schooling get in the way of
sd...@cornell.edu my education" -Mark Twain
---------------------------------------------------------------------
Michael D. Bayne
: Distributed Shared Memory is actually a hot topic in OS's these days.
: Its a neat communications paradigm, but still needs research work.
So how does one get around the fact that in order to access *anything* that's
in memory on some other machine you have to copy it (over the network) into
this machines address space. That seems like too large a performance
hit. Harddrives are faster than networks. The only way around copying the
memory is to have all the processors on the same bus and then that's not a
distributed system, that's a parallel system. Investigation in sharing memory
under 'that' paradigm is indeed a hot topic in OS design today.
Stephen D. Holland
I agree that DSM is slow, but that doesn't detract from its potential
usefulness. Consider 20 machines in a network, not necessarily physically
near each other, sharing a "scratchpad" for communications. Don't come up
with a list of limitations, we all know they're there. That's why the research
work is needed to turn this into something genuinely useful
-Steve
--
-----------------------------------------------------------------------
Steve Holland N1RAO "Opera is when a guy gets stabbed in the back and
sd...@cornell.edu instead of bleeding, he sings." -Ed Gardner
-----------------------------------------------------------------------
Michael D. Bayne
: usefulness. Consider 20 machines in a network, not necessarily physically
: near each other, sharing a "scratchpad" for communications. Don't come up
: with a list of limitations, we all know they're there. That's why the
: research work is needed to turn this into something genuinely useful
I agree that what you describe is a useful scenario, however, you don't need
distributed shared memory to do that. I would say that's overkill. Let's look
at something like virtual memory. Before it existed, one could swap things to
disk themselves, however it turned out to be very useful to have this built in
to an operating system so that it could be done efficiently and programmers
didn't have to worry about doing it themselves. In the case of distributed
shared memory, it seems to have very limited application and thus I suspect
you'll find that people will end up "doing it themselves" and not having it
built into an operating system. The overall performance decrease from having
DSM built in to an OS isn't outweighed by it's fringe usefulness.
Manuel Lemos
: If you have a Ethernet connection or any fast network connection to
: a UNIX system, the following may not be such a bad idea:
: Run a server on your UNIX account that malloc() UNIX system memory
: and allow a client running on AmiTCP side to use those memory over the
: network!
: If this sounds like parasite, you're right :-) But it could be
: one possible alternative to virtual memory systems. Simply write a
: server that process memory requests, malloc() a block of memory
: from the UNIX system, and let the client program read/write this
: piece of memory. A client program on AmiTCP side will do what most
: virtual memory programs do except that it gets its memory resource
: from the server across the network instead of a cache on disk.
: Possible? Implementable? I think so :-)
Sorry, too inneficient to be considered. You need at least a MMUable Amiga
and even with 10Mbps DMAable Ethernet card, the system would crawl.