[comp.arch] Loose vs Tight coupling for OLTP

dhepner@hpisod2.HP.COM (Dan Hepner) (03/17/90)

From: shekita@provolone.cs.wisc.edu (E Shekita)

>[ .. due to concerns of bus saturation ..] Over the long run, I think 
>shared-nothing configurations (like Tandem offers) will prevail for 
>transaction processing systems.

Since Tandem distributes the database over its processors, it seems
a little exaggerated to describe them as "shared-nothing".  Substantial
communication requirements exist between those processors to effect
not only remote disk IO but also concurrency control and distributed 
transaction commit.  Shared memory systems have a similar problem to 
solve, of course, but it's reduced by the lack of a remote disk access
requirement. All interesting multi-processor OLTP systems end up with 
shared concurrency and transaction control.

Thus for Shekita's claim to be true, one must believe either that:

1) The communication requirements to achieve multi-processor IO, 
   concurrency, and transactions are low enough to not dictate the 
   decision.
or

2) Using external communication paths results in superior overall
   performance regardless of the communication requirement.

How likely is either to be true?

Some other possibilities:

1) Being bus bandwidth bound is no more difficult to address than other
   potential bottlenecks in an OLTP system.

2) Alternatives to the bus bottleneck are no less likely to themselves
   be bottlenecks.

Dan Hepner