We have a Dell Equalogic iSCSI SAN. We have been having latency issues with it and Dell gave us their best practices for Nexus 5k config to implement. Prior to the change we had MTU Jumbo settings set system wide and pause no-drop was not configured. The only traffic on these 5k switches is iSCSI traffic. There is nothing else so I did not feel the need to use any QoS settings of any kind. We had flow control send and receive turned on on each interface as well. Below is the Dell best practices config I implemented. Ever since implemented I have been seeing millions and billions of TxPause frames counting up on many interfaces. Does anyone have any ideas?
##2.10.1 Configure CoS for iSCSI
interface ethernet 1/1-40
untagged cos 4
##2.10.2 Define a QoS map for iSCSI
class-map type qos class-iscsi
match cos 4
##2.10.3 Define a QoS policy map for iSCSI
policy-map type qos policy-qos
class type qos class-iscsi
set qos-group 3
##2.10.4 Define a Network QoS class map
class-map type network-qos class-iscsi
match qos-group 3
##2.10.5 Define a no-drop policy map and enable jumbo frames
After some research I have answered my own question. By enabling pause no-drop it has sliced my available buffer from 470KB to 90KB. It now sends TX pause frames out at 37KB and Resume frames at 17KB. It will not use available buffer space in excess of 90.24KB It appears I will need to use a command similar to "pause no-drop buffer-size 442000 pause-threshold 300560 resume-threshold 243100”. My problem now is, what to set all those values to. I wonder what percentage of buffer usage I should have it send pause/resume frames.
on our 5596, we worked with Cisco TAC to configure buffers to nearly 400k. The manual is wrong or ambigous and TAC asked why we had them set so low while troubleshooting the issue. The PFC on the port-channels was set to off and that seemed to resolve our issues. PAuse frames are now almost entirely absent.