| T.R | Title | User | Personal
Name | Date | Lines |
|---|
| 3072.1 | You can use cluster aliases, but at a price | STAR::VATNE | Peter Vatne, VMS Development | Thu Jul 12 1990 18:57 | 22 |
| It turns out that cluster aliases can be used with DECwindows, but
at a fairly hefty performance cost. The object name you are looking
for is X$X0. Therefore, the command you need is something like:
NCP> SET OBJECT X$X0 ALIAS OUTGOING ENABLED
If you do this, then any connection originating from any node in
the cluster will identify itself as being from the cluster, and
not from an individual node. This simplifies the security setup.
However, events and replies returned from the server get sent to
the cluster router, which in turn forwards the packets to the
originating node. This results in significant overhead.
Since we don't document how cluster aliases can be enabled, the
statement in the VMS DECwindows User's Guide that the nodename
must be an individual node name and not a cluster alias still stands.
If you do something unsupported like setting the cluster alias, then
of course you can expect the behavior to not match the documentation.
I suspect what has happened is that someone enabled the alias on
MSBCS, but not on CHIRPA. Go ahead and set the alias, and let us
know if the performance is acceptable to you.
|
| 3072.2 | I really hope I'm wrong about this... | IO::MCCARTNEY | James T. McCartney III - DTN 381-2244 ZK02-2/N24 | Fri Jul 13 1990 11:22 | 8 |
|
But doesn't enabling alias effectively double the network traffic
per event? Sounds like a really high price. In an environment like
most of Digital's larger sites -- where we're already running out
of time on the wire -- this doesn't seem worth it. Especially if
all it "saves" is having to type a few additional entries in a table.
James
|
| 3072.3 | | MSBCS::SKI | | Fri Jul 13 1990 13:19 | 17 |
| >> But doesn't enabling alias effectively double the network traffic
>> per event?
I must admit I do not know; even our local DataComm guy doesn't
know. Is there anyone out there who can give a definitive answer
on this?
>> this doesn't seem worth it. Especially if
>> all it "saves" is having to type a few additional entries in a table.
Agreed. If this turns out to be the case, then I certainly won't
enable the aliases.
So far I have not noticed a difference in response times between
the cluster using the alias and the cluster which is not.
However, I might just be getting lucky, or perhaps the part of
the network I am in isn't too heavily used.
|
| 3072.4 | You've hit it right on the head | STAR::VATNE | Peter Vatne, VMS Development | Fri Jul 13 1990 13:21 | 8 |
| >But doesn't enabling alias effectively double the network traffic
>per event? Sounds like a really high price. In an environment like
>most of Digital's larger sites -- where we're already running out
>of time on the wire -- this doesn't seem worth it. Especially if
>all it "saves" is having to type a few additional entries in a table.
This is exactly my point about "a fairly hefty performance cost",
and the main reason we don't document cluster alias support.
|
| 3072.5 | That settles it | MSBCS::SKI | | Fri Jul 13 1990 13:26 | 3 |
| Well then, I'm off to kill the alias.
Thanks muchly!
|
| 3072.6 | Alias performance tutorial | STAR::BECK | Paul Beck - VMS Development | Tue Aug 07 1990 14:44 | 21 |
| As a postscript to this discussion (I haven't followed this conference for a
long time), a word from the horse's mouth (the designer of cluster alias):
There are two reasons why performance using an alias is degraded relative to
using a specific node name. First is routing: every packet to or from the alias
must pass through at least one router (in the cluster) and sometimes two
(add the designated router if the packet originates off-LAN). This means that
the packet appears on the Ethernet 2-3 times.
The second factor is packet size: if both sender and receiver are on the same
Ethernet, DECnet will optimize the buffer size used from the default 576 to
1498. When an alias is in use, one of the nodes is off-LAN (the alias node,
which is not a real node). We can't fake this, because the same bit that tells
DECnet to optimize the buffer size is used to tell on-LAN endnodes to bypass
the router and address the node directly. Since there is no node on the Ethernet
with the alias' hardware address, this would prevent communications.
So with an alias, smaller buffer sizes are used, tripling the number of [large]
packets in use. (This effect is primarily of interest for traffic like file
transfers; interactive traffic which doesn't involve packets larger than 576
bytes would not see this effect).
|