Request for Comments

[Douglas Eadline]

On Sun, 23 Jan 2000, Michael Croft-White wrote:

> 
> I don't know if this is even feasible or practical at the moment but I
> just wanted to find out what other people think about a little idea I
> had and if they can forsee any problems I haven't though of yet.
> Anyway, the idea is based upon making a beowulf system seamlessly
> integrated to ease programming and make life a little easier.

Ah, the Holy Grail of parallel computing. A noble quest.
> 
> The idea is basically that a piece of code sits between the kernel
> base and the SMP layer, which extends the bus across a network,
> resulting in what appears to be a n-way SMP machine instead of n
> individual machines working on the same thing. This would allow
> programming for the system to just require multi-threading instead of
> using additional libraries or whatever. Also it would make the beowulf
> architecture open to a few more people, e.g. the person with say four
> or five machines lying around who wants just a powerful workstation.
> 
Well, in principal this is good idea. In practice, it is very difficult
to do this and get the performance you need without resorting to 
special hardware and software.  This idea is basically what NUMA (Non
Uniform Memory Access) architectures address.

Trying to implement such a memory model on Beowulf presents the
following problems:

1) cache coherence (keeping track of dirty memory)
2) optimizing "non-uniform" memory access is very difficult
   (you need to consider vast differences between accessing
    local vs. non-local memory)

Simply put, no matter what "programming API" you put on top of 
a Beowulf, you still are communicating with messages (copying 
data) at rates from 10-100MB/sec and latencies of 20-200 us.

There have been several "shared memory" on top of clusters
packages written, but the are not a universally useful solution.
(i.e. they work well for some problems but not for all)

If you want the details, check out George Pfisters book
"In search of Clusters". It provides excellent (and detailed) analysis
of various parallel computing designs.

Doug
-------------------------------------------------------------------
Paralogic, Inc.           |     PEAK     |      Voice:+610.861.6960
115 Research Drive        |   PARALLEL   |        Fax:+610.861.8247
Bethlehem, PA 18017 USA   |  PERFORMANCE |    http://www.plogic.com
-------------------------------------------------------------------


-------------------------------------------------------------------
To unsubscribe send a message body containing "unsubscribe"
to beowulf-request@beowulf.org