Point at which WAN latency affects perceived app response?

Does anyone have any experience with the effect of WAN latency on
applications accessing an Oracle database over a WAN?

We are using Oracle Server 9ir2 (9.2.0.7.0) as our back end database
for a packaged 2-tier client/server business application. The
development tool and network APIs for the application are fairly
obscure, but they must touch on SQL*Net eventually as the Oracle client
is required.

We have two sites. Each has an Oracle database for the portion of the
app that serves that site's business function. But of course, people
from A need to access the primary application for B, and people from B
need to access the primary application for A.

To this point we have served the cross-site requirements by installing
both sets of software at both sites, and simply accessing the Oracle
data directly across the WAN. Brutal and perhaps not elegent, but
simple and straightforward (particularly compared to an experiment with
Citrix).

We are now in the process of replacing our T1 leased line between the
sites with an MPLS-based network. Each primary site A and B has a 3
Mbps MPLS port. There are a few other minor sites that connect, none
that run the business apps.

Bandwidth on the MPLS connection is good: transferring large files of
random numbers I see close to full port thruput between the sites.
E-mail, web, etc work fine.

But when we attempted to bring up the Oracle-based apps the results
were not so good. Perceived performance got worse by a factor of 4,
and stopwatch timings confirm that the new configuration is slower by
at least a factor of 3.

Looking at ping times, I find the following for the two different
connections:

Packet size Leased Line MPLS
=========== =========== =========
256 6 ms 48 ms
512 11 ms 55 ms
1024 17 ms 69 ms
2048 26 ms 84 ms

Now, clearly less latency is better than more latency. Still, 60 ms
latency between sites is considered quite good on most WANs of any size
and I would think most Oracle users would be in that range or higher.
Would 60 ms be sufficient to cause the application to appear to the end
users as noticably slow?

I have been trying to watch the connection with Ethereal, and one thing
I noticed was that there seem to be a lot of fragemented packets in the
100-300 byte range. Other tools confirm that the non-fragmented MTU
for the link is 1500 bytes (same as the leased line). Does it make
sense that either the server or the router would break larger packets
up that small?

My knowledge in this area is like the Platte River (wide and shallow),
so any input would be appreciated.

sPh

.

Relevant Pages