annuler
Affichage des résultats de 
Rechercher plutôt 
Vouliez-vous dire : 
cancel
494
Visites
0
Compliment
2
Réponses

cat9500 qos queue-softmax-multiplier

damien-gluski
Level 1
Level 1

In order to reduce packet drop on catalyst 9500-32C and 9500-48YC, I configured the qos queue-softmax-multiplier setting. Tricky part was to figure out which value is recommended.
I did several changes and monitor the metric "Drop queu full" in snmp :
From qos multiplier value 0 to 2400 : we notice huge improvement and less packet drop due to queu full
Form qos multiplier value 2400 to 4800 : result are not flagrant, no real improvement noticed.
My concern/question is about the maximum value 4800 : is it recommended to configure it ?
Pro for 4800 :  it provide the maximum capacity to absorb TCP burst and reduce packet drop 
Con for 4800 : as explained on some cisco documentations,  higher the qos multiplier is, less chance to handle correctly concurrent TCP burst.   Logically (but not written by cisco) with higher qos multiplier value and during concurrent TCP burst, an interface can take majority of shared available buffers and less buffer remains for the other interfaces.  Expected result could be  high speed transfert for some session and really bad speed performances for the others sessions. 
But does DTS (Dynamic Threshold Scale) shouldnot handle high shared buffer utilization ? If the shared buffer utilization reach like 75%,  the softmax of interfaces should be reduced automatically. So the first interface concerned by TCP burst should not be able to take majority (80%) of the shared buffer if other interfaces need it. the DTS should reduce it. In this case no more question reagarding the qos multiplier value, it must be configured to the maximum value.
Until this question is not  clarify, I configured on my cisco cat9500 (count 50)  the qos queue-softmax-multiplier 2400, by precaution principle in case of concurrent TCP burst.
Thanks for your feedback and expertise on this parameter. 

2 RÉPONSES 2

qos queue-softmax-multiplier 1200 <<- 1200 is test and it work for c9k series 

MHM

On the 9500 Series high-performance (ex 9500-32C) , the configurable qos value range is between 100 and 4800.
The 1200 is the maximum for lower model but not for 9500 Series high-performance
I have already tested 1200, 2400 and 4800.
All of them works actually, but my concern is about potential side-effect of configuring high value like 4800.
Does it could cause more packet drop during concurrent micro burst ?
Regarding the value 1200, from my tests it is not enough to absorb micro burst on 9500 Series high-performance (ex 9500-32C) .

Regards