Is QOS over the shared bus possible?

[Geatz]

We have a 6509 with 6148a Gigabit line cards. We are noticing latency on servers and high amounts of PAUSE outputs erros from netflow enabled ports. I know these line cards are not meant for servers and that getting server gig line cards would probably fix our problem. However, this may not be an option, so in the meantime I'm wondering if there are any QOS options I can use to prioritize layer 2 traffic - COS - through the shared bus for a few of our more important servers?

I have tried configuring "mls qos trust cos" and "mls qos cos 5" on the interfaces but that hasn't seemed to have any affect on the amount of PAUSE outputs or the latency.

 

[unclerico]

this is what you get with oversubscribed hardware. i don't believe there is any way to get around it without going to non-blocking/line rate cards.

I hate all Uppercase... I don't want my groups to seem angry at me all the time! =)
- ColdFlame (vbscript forum)

 

[VinceWhirlwind]

I think it isn't accurate to call them "Pause errors" - they're not errors, they are flow control messages.

Also, a high Pause count could simply indicate that the servers are not participating in flow control, meaning the switch is sending and resending messages the server might be ignoring.

So, assuming the problem is caused by the servers trying to send more data than the switch can pass, what you could do is this: the oversubscription on the switch is done in 6 blocks of 8 ports. So, monitor your switch ports for a while, and then repatch them so that the 6 busiest ports are each in a separate block on the switchport module, and the next 6 busiest, and so on.

 

[Geatz]

You're right, I probably shouldn't be calling them errors. The real issue is that I'm seeing latency; however, the only abnormal effects I've seen are these pause outputs.

Let e start by shedding a little more light on what I am doing here. The problem ports are connected to our compellent san, for trouble shooting purposes I am specifically looking at one transmit port and one input port on the SAN vlan. These two ports are in direct communication with eachother, the server port outputs the frames to the compellent san port. The Input port/compellent san port is where I am seeing the pause outputs.

I'm also fairly certain the compellent san port is netflow enabled for a few reasons, first, the port is set to dynamic and configuring itself and second, the vendor told me it was. However, not being the trusting type, I have tried disabling netflow on the port, but I just got a lot of dropped packets.

I have already isolated these two server ports, they are the only ports utilized on a set of 8. When isolation didn't solve the problem I concluded the problem was either the shared bus or the servers themselves. I've pretty much made the assumption that the servers should be able to do line rate and so I haven't begun to put the blame on the servers to much yet.

We also have 5 of these cards in our 6509 so I assume that means we have a possible 25 seperate dataflows that will be sharing that one 32Gb connection. We are definetly not hitting 32Gb yet, but being able to service only one data flow at a time is where I believe we are hitting our bottleneck.

So I guess my question still remains, is it possible to do anytype of qos over the shared bus? Does the shared bus have prioity queuing or is everything just round robin?

 

[brianinms]

Are you confusing the terms flow control and netflow?