Enterprise Beowulf and Beowulf on a Chip

Mark Hahn hahn at physics.mcmaster.ca
Wed Jun 18 23:03:09 EDT 2003

> In my own search for finding out whether Beowulf makes sense for
> enterprise computing (specifically on-line transaction processing
> (OLTP)),

OLTP has significant throughput-type parallelism, so can be done
by clusters quite nicely.  it could be that you'd want a low-latency
interconnect for lockmanager-type synchronization, but wouldn't 
necessarily need much bandwidth.

> This work emphasizes "large memory stall times" as being
> a primary culprit limiting OLTP performance.  The solution

right: OLTP is computationally trivial.  so you can either use 
CMP-type parallelism (where stalls still waste the CPU for the 
duration, but CPUs are cheap) or SMT (which they reject for a 
practical reason, namely that CMP is a lot easier to throw together.)

> multiprocessing (CMP).  This makes me wonder if OLTP
> can benefit from simply lots of processors (with fast interconnect)
> to utilize more L1, L2 cache simultaneously (a might bit larger
> than a chip I might add!!)...

it's not clear how much good cache does you for OLTP - 
if you look at figure 6, you see that miss rate decreases 
between P4 and P8 - that is, when you have 8 1M caches, you
start to approach a working set.  

in figure 8, you see that misses are still something like 33% of 
execution time, though, so even 12M cache is not terribly effective.

>  If this approach makes sense for OLTP doesn't a Beowulf
> make sense for OLTP work now?

clusters are great for OLTP, but Beowulf is usually considered a fairly
specific kind of cluster, which is tuned in ways that are not much of 
advantage for OLTP.

> If Beowulf makes sense on the macro level does it make
> sense in the micro-level or perhaps in the fractal sense of
> a self similar architecture (exploiting even more hierarchy)?

I do not believe there is any problem obtaining whatever OLTP performance
you need.  do you need 700K tpmC?  I really don't think so.  making OLTP 
cheap (even at modest performance) is an entirely different topic.

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