Very Computer

We are running a 6520 in an LAVC cluster with 1 3400 workstation
and a 3400 that runs misc applications.

I'm looking for a ballpark estimage of how much performance
we could gain on the 6520 by running it unclusterd.

Thanks.

Jim
+------------------------------------+--------------------------------------+

|  VeriFone Inc.                     |   uunet!verifone!jimmy_t  (UUCP)     |
|  100 Kahelu Avenue                 |   808-623-2911            (Phone)    |
|  Mililani, HI 96789                |   808-625-3201            (FAX)      |
+------------------------------------+--------------------------------------+

Depends on how much locking is going on between the nodes (MONITOR DLOCK) and
how much MSCP serving the 6520 is being asked to do by the Ethernet-connected
nodes (MONITOR MSCP).  Are the MicroVAXes booting from the 6500, or do they
have their own system disks?  Are the MicroVAXes paging to their own disks, or
paging to disks on the 6500?

-Dan Wing, Systems Administrator, University Hospital, Denver

--

Well, if you're not loading the cluster (and the MSCP) stuff, you will gain
some resources on you're system. It might give some of your applications
some extra air.

You're distributed lock manager will be less busy.

No diskserving by the 6520 for the VAX3xxx will decrease the load on your
diskchannels.

And so on, and so on........ (fill in the ret yourself).

So, YES you might benefit, BUT to what extend ???? (don't know what you're
running on the 6520, and how much the VAX3xxx systems use the 6520 as a
server).

Cheers, Hans

--->>>  Happiness is just an illusion filled with sadness and confusion.  <<<---

        ITT Gouden Gids                        
        Data Center                     Phone: +31 20 5676797
        Hoekenrode 1                    Fax  : +31 20 6910374
        1102 BR  AMSTERDAM
        Netherlands

        Disclaimer: No responsibility is accepted by anyone on anything
                    I say, write, or do. This also includes myself.

--->>>   It's not reality that's important, but how you perceive things   <<<---

--
James,

Your question concerning performance improvements when systems
are no longer clustered is a good one. Unfortunately, there is no
easy way to predict the performance of a system when you remove a
system from the cluster.

The actual performance will vary depending upon may factors. Some
of these factors include locking activity, IO activity, and many
others. Without detailed knowledge of your particular
applications environment, any answer would be, at best, a guess,
and at worst, could be flat wrong.

I hope that the above is helpful. If I can be of further
assistance, please feel free to contact me via Email or phone.

- Bob
+--------------------------------------------------------------------------+

| Robert Gezelter Software Consultant         Voice:   +1 718 463 1079     |
| 35-20 167th Street, Suite 215               Fax:       (on Request)      |
| Flushing, New York  11358-1731                                           |
| United States of America                                                 |
+--------------------------------------------------------------------------+

Quote:> We are running a 6520 in an LAVC cluster with 1 3400 workstation
> and a 3400 that runs misc applications.

> I'm looking for a ballpark estimage of how much performance
> we could gain on the 6520 by running it unclusterd.

The MSCP server runs on the interrupt stack. The server function of the
Distributed Lock Manager, I believe, also runs on the interrupt stack.

So an upper limit of how much of the 6520's CPU is being tied up serving
the other nodes is how much time it's spending on the interrupt stack. Of
course there are other resaons for it to spend time there. MONITOR MODES
shows this.

Another issue is whether the 6520 is being hampered by requesting locks
that are mastered remotely (by the 3400s) as opposed to locally. You can
look at this through MONITOR DLOCK.

This hits you in two ways. Communication with a remote lock master takes
both extra CPU cycles, and introduces a delay due to the communication path.
The extra CPU cycles will most likely show up as Kernel Mode and Interrupt
Stack time. The communication delays aren't so easy to measure, but an
ethernet doesn't provide much bandwidth for systems like these.

I'm inclined to think that if you're thinking of running your systems
unclustered (or moving the 6520 out of the cluster), you probably have the
resources the 3400s need local to them. On this basis, the 6520 is not
going to be doing much work MSCP serving. You would also hope that the
algorithm used for deciding which node masters locks would be bright enough
to make sure that it wasn't the 6520 mastering the locks mainly accessed by
the 3400s and that the 3400s weren't mastering the locks mainly accessed by
the 6520. [If anyone knows how this works, I'd love to hear]. (Of course,
you can mess this up buy your setting of LOCKDIRWT). So you might not be
seeing too much of an impact by having the 6520 clustered.

Mark

Dept Mech Eng, University College, London. UK

                        Read at your own risk.