| Well, IMPO, I agree that the DLM is not designed for high transaction
throughput. However, it is designed for high availability. This is based on
the fact, that you get less tps from cluster in terms of percentage, then you
do from a single node. Why do you think that our TPC-* tests are run with
partitions of the database all being accessed from a single node?
Also, why do you think that we spend so much effort in getting the numbers of
locks down. It is not because lock mgmt is so fast.
Now, there certainly can be ways to make the DLM even slower. my guess is that
O could not figure out how to optimze. Or maybe they figured that if they
built it a different way and only use the DLM for certain things, that they
would have something to tell the press like DEC stuff is too slow, that is why
rdb is slow and anyone else who uses it is slow.
chuck
|
| FWIW, I think that Oracle's marketing only supports their (probable) goals.
If O had used Digital's Distributed Lock Manager, then they would be
implicitly condoning it and they would ruin one of their marketing
strategies against Rdb. (i.e. Customer to O: If DLM is so bad, then
why are you using it?) I know that they use DLM for some things, but
they seem to denigrate it to a minor role.
In addition, the VAX Distributed Lock Manager is not portable. If O
creates their own "DLM", then they can use it on other platforms. If
you accept that logic, then given that they've gone to all of the
trouble of creating their own DLM, why would they then admit that
Digital's is any good? Once again, if they did admit that it was OK,
then they would be implicitly admitting that Rdb (based on LM and by
extension DLM) must be OK. Oracle is certainly not known for inept
marketing.
BTW, when I took my VAXcluster Distributed Lock Manager course
(admittedly, 5 years ago) there was a distinction made between local
locks and distributed locks. If a resource was local, then so were its
locks. If a resource was distributed, then so were its locals. At the
time, there was an order of magnitude difference in performance between
a local lock (contained within 1 CPU) and a distributed lock (which had
to travel across the CI which is slow compared to memory.)
I would assume that Oracle would have the same problems and issues to
solve as we did in designing a DLM and so I can't believe that their
DLM's performance would be better than ours. After all, we've had a
few years to refine the design. Nonetheless, the goal of both database
designs has to be to reduce the number of remote locks. (The CI is the
same speed for both O and DEC, assuming that they use the CI.)
|