Conference decwet::winnt-clusters

814.0. "I have an answer.." by TAVEIS::SHALOM () Sun May 18 1997 06:28

Guys,

	The colour is returning to our cheeks. Technical discusiion!!!

	Some points - maybe worth making.

	re .29 , Carl's suggestion to replace a Quad Pentium with a Quad
Rawhide to gain performance. Carl, without being nasty - Can you provide a set
of numbers that will show that ANYTHING besides marginal improvement is
achievable, something on par with the Huge price differential??? I need this
figures, I could start selling 4100's today to replace single-node Saturated
Pentium Pro's.

	re .31 - I love the reference to "lightweight" LM (as opposed to DLM)
being built from simple mutexes and semaphors. As a VMS Veteran, I sugggest
looking at the data structures involved in the Lock manager - the Lock blocks,
the Resource blocks, the hashtables, the Deadlock detection algorithm. On the
other hand, "porting" VMS's DLM is not a good idea, but copying the design
should shorten the time to market very significantly. Just pick up your old
copy of Ruth Goldenberg's book...

	Finally: ORACLE is promising to do Parallel Queries on a cluster using
OPS (Oracle Parallel Server). THis is decision support, --BY FAR-- heavier than
transaction processing where you do point retrievals and updates, not massive
walkthrough of GIgabytes of data (select * from employees,products,sales,
regions, pricelists where ... and ....  ). There, the ability to split a single
query and have it ready in less than a week is CRUCIAL. is BIG TIME. WHen
Oracle has that - and they are going to do that, easily - port DLM, use Raw
partitions to prevent NTFS from messing up with the data-

	- and when Digital and others supply Memory channel access to
coordinate this stuff -

	You'll see Microsoft change religion. Don't be surprised if NT5 is
delayed to allow the SQL people to squeeze some cooperative processing into a
cluster environment...

		... and have a nice life.

						Shalom
and ... and ...