Performance problem with Oracle app

[pat2006]

Hi.

Here is the problem.

Our existing LAN is flat - one IP subnet, with over 1000 nodes. There are Accelars at the core but the PCs connect via 10BaseT hubs.

We are in the process of upgrading to a segmented network using 8600's and switches (100Mbps) for clients. This new net is connected to the old net since all apps and servers are still on it.

The new net uses different IP subnets to segment the traffic.

The performance of the new network is excellent (at least I think so). However, we have an Oracle app that seems to behave opposite to what you expect. Let me explain.

When the client is on the old net the performance is better than if you move the client to the new net (a different IP network basically). Rememeber, the servers are still (unfortunately) on the old network.

In talking to an Oracle expert, he seems to suggest that the database was poorly designed.

I believe that for some reason it is not optimised to work across subnets.

I don't think any "tuning" of the database was ever done.

Can anyone out there shed some light on this?

It's really disappointing to have a new faster network and this particular application's performance has gotten worse.

I know it's a long post, but thanks in advance.

 

[anthonyanderberg]

Its possible the database was poorly designed, but its hard to imagine a design that would affect one group of clients and not the other - or one subnet but not another.

How are the old and new networks linked together? Perhaps the real problem is that the database is returning lots of (un-needed) rows and a bottleneck between the networks is making it seem slow. How do other applications/file sharing perform?

The only other thing that occurs to me is that in your new network you'll need to be aware of the speed/duplex settings of your switches and PC NICs. Duplex mismatches are still (unfortunately) a common problem that results in weird slowness, esp in large data transfers or other 'chatty' situations. If you've got modern PCs with up-to-date drivers setting the speed/duplex to autonegotiate should work, but its worth double-checking.

 

[pat2006]

Thanks Anthony. Try telling a DBA that their database is poorly designed!

The 2 networks are linked via a 1Gbps FO link between the 8600 and an Accelar (They are physically located next to each other).

As for other apps? File sharing is ligthning fast between networks. I haven't heard any complaints about email and SAP.

The comments about duplex mismatches is interesting. All of our clients are either Win2K or XP. The PCs are quite new - Dells, HPs etc.

We use 460's to connect clients. Can the 460's give an indication of when a duplex mismatch occurs?

Thanks again.

 

[anthonyanderberg]

If file sharing is quick then duplex mismatch or bad cables seem less likely (100Mbps isn't as forgiving as 10Mbps)... but for what its worth - depending on which side is half duplex your 460s port statistics might show collisions and FCS/CRC errors. If the 460s port setting show "Auto-100/Half" that would be a good indication that the PCs might be manually set and the switch is getting the speed right but is having to assume half duplex.

What happens if you share files from the database server to the clients? - that would really help you narrow things down.

 

[Kwazzy]

If I may chime in here, and I believe this is where Anthony is heading..

In a mis-matched duplex setting, if one side is hard set to 100 full and the autonegotiation enabled on the switch (or vice versa), you more than likely will see the Autonegotiation end at 100 half. It will detect the speed in auto, but no duplex. Both sides must have the autonegotiation signal in order for both ends to setup full duplex. Otherwise, it will default to the lowest denomiatior, half duplex. So your best be is to either Statically assign the duplex on both ends, or have autonegotiation enabled on both ends.

Leaving the link mis-matched will cause you some major issues since you are retransmitting about 75% of you traffic, and your data transfers will crawl. As Anthony has described above you will see these collisions and FCS/CRC errors.

A good indication if you cannot view the errors on a link... Autonegotion Enabled and you have a link at 100-Half Duplex...

Funny note on this... I took a Cisco class about three years ago, I actually had to argue this point to the instructor. What blew my mind on this, it was a dead end aguement with him because the class was hardware less (Virtual Labs Scripted to Emulate the work), so I just informed those from my company the truth. Unfortunate that he calls himself a consultant and instructor. Need less to say the class was a waste, the documentation was good.