Thanks for the feedback ...

No problem. The general issue I am seeing is that network engineers across the industry are familiar with traditional store and forward switching. The concepts of cut-through switching are more closely aligned with how fibre channel operates. Therefore what you end up with is sub-optimal switch configurations for carry iSCSI traffic; buffers are over-run and therefore you get retransmit on the SAN's. Whilst TCP will cope with this due to guaranteed delivery it affects performance. Specifically some topics that could be useful to be covered are:
1) Buffer sizing considerations for edge ports and inter-switch port channels
2) Ethernet Flow Control (and whether it should be used)
3) PFC and support end-to-end, i.e hypervisor/OS and SAN inclusive.
4) QOS/COS tags, QOS configurations pre-DCB, channel configurations with DCB
5) Delayed Ack and best practice. Where it helps, where it doesn't.
6) Retransmit tolerance, at what point is TCP retransmit a problem.
7) Spanning-tree free network design
8) Layer-3 iSCSI considerations (most say don't do it, but this *may* change with layer 3 switching functionality coming into cut-through switches)
9) End to end testing and load generation.

Happy to discuss further if you like. Flick me an email (presuming you can see the address I have signed up with)

The SIS team is actively engaged in various test/documentation projects that are focused on the topics you list above. I'm forwarding your feedback into the team. BTW - your email address is hidden by default in your profile. Please flick an email my direction so we can connect: chris_almond AT dell dot com

Thanks - Chris

Jeremy - have you had a look at our EqualLogic Configuration Guide yet? (linked from this page). Its charter is to be that "general guide" that you ask about in your original comment. We have it on a quarterly update cycle. Many of the implementation topics you ask about will eventually make it into that guide.

Thanks again for the great feedback.
- Chris

"No problem. The general issue I am seeing is that network engineers across the industry are familiar with traditional store and forward switching. The concepts of cut-through switching are more closely aligned with how fibre channel operates. Therefore what you end up with is sub-optimal switch configurations for carry iSCSI traffic; buffers are over-run and therefore you get retransmit on the SAN's. Whilst TCP will cope with this due to guaranteed delivery it affects performance. Specifically some topics that could be useful to be covered are:
1) Buffer sizing considerations for edge ports and inter-switch port channels
2) Ethernet Flow Control (and whether it should be used)
3) PFC and support end-to-end, i.e hypervisor/OS and SAN inclusive.
4) QOS/COS tags, QOS configurations pre-DCB, channel configurations with DCB
5) Delayed Ack and best practice. Where it helps, where it doesn't.
6) Retransmit tolerance, at what point is TCP retransmit a problem.
7) Spanning-tree free network design
8) Layer-3 iSCSI considerations (most say don't do it, but this *may* change with layer 3 switching functionality coming into cut-through switches)
9) End to end testing and load generation.

Happy to discuss further if you like. Flick me an email (presuming you can see the address I have signed up with)
"

Excellent points... I have gone back to unmanaged switches for iSCSI. Every single thing we tried on managed switched seem to degrade performance. We run really low end iSCSI sans though.

Hi Chris,

Yes, I have read the configuration guide. It has some really useful information in it. As it is generic however, it doesn't provide information in context with other vendors best practice guides. For example, it talks about buffer sizes on switches; it would be really useful if switching vendors then produced implementation/best practice guides to verify that those buffers are available, and assigned to the correct switch port (in some vendor products large buffers are only assigned to FCoE traffic), and then how to verify that you aren't getting buffer overruns. What I find right now it that I may read a Dell Configuration Guide for the EqualLogic SAN's which states that something should be enabled at a switch layer or host layer, but nothing in the switching or host vendors configuration guides that are authoritative on the topic.