Confusing with Class Based Weighted Random Early Detection

[ctranminh]

I have a project with QoS between 2 drilling rig. I worked for a time to have this below configuration:

class-map match-any BULK_DATA

match ip dscp af33

match protocol syslog

match protocol snmp

match protocol icmp

match protocol ssh

match protocol telnet

class-map match-any BUSINESS_TRAFFIC

match ip dscp cs3

class-map match-any VIDEO_CAMERA

match ip dscp af41

match ip dscp af42

match ip dscp af43

class-map match-any MUST_ONLINE

match ip dscp cs6

class-map match-any REAL_TIME

match ip dscp ef

class-map match-any BEST_EFFORT

match ip dscp af11

match ip dscp af12

match ip dscp af13

policy-map OUTPUT_POLICY

class MUST_ONLINE

bandwidth percent 12

class REAL_TIME

priority percent 25

class VIDEO_CAMERA

priority percent 15

class BUSINESS_TRAFFIC

bandwidth percent 4

random-detect

random-detect precedence 3   32    40

class BULK_DATA

bandwidth percent 15

random-detect dscp-based

random-detect dscp 0   30    38 !(When I show policy-map I see all of management traffic (ssh, syslog, snmp) fall in default class, therefore I changed

                                                settings in order to make them less drop probility than af33 traffic when congestion occur)

random-detect dscp af33   28    38                            

class BEST_EFFORT

bandwidth percent 19

random-detect dscp-based

random-detect dscp af11   24    36  

random-detect dscp af12   22    36

random-detect dscp af13   20    36

class class-default

fair-queue

My question is "Do I need change the default minimum threshold and maximum threshold of DSCP based WRED?"

I want to clearly differentiate that BULK_DATA traffic is more likely drop than BUSINESS_TRAFFIC and less likely to drop than BEST_EFFORT traffic when congesting. That why I change some default settings. 32>30>28>24>22>20

But Cisco say the Default settings are, and recommend you should not change these

                                                                   Min Thres    Max Thres

   af11       0/0               0/0              0/0          32                40      1/10

   af12       0/0               0/0              0/0          28                40      1/10

   af13       0/0               0/0              0/0          24                40      1/10

   af21       0/0               0/0              0/0           32                40      1/10

   af22       0/0               0/0              0/0           28                40      1/10

   af23       0/0               0/0              0/0          24                40       1/10

   af31       0/0               0/0              0/0          32                40       1/10

   af32       0/0               0/0              0/0          28                40       1/10

   af33       0/0               0/0              0/0           24                40       1/10

   af41       0/0               0/0              0/0           32                40       1/10

   af42       0/0               0/0              0/0           28                40       1/10

   af43       0/0               0/0              0/0          24                40       1/10

Therefore, I think there is no difference between aff11, aff21, af31, and af41, for example

I misunderstand, Plz help me!!!

[smehrnia · Accepted solution]

Hi,

whenever u want to change the default values for queue thresholds and buffers, u have to have pretty good knowledge of QoS behaviour and your device capabilities! if you dont feel that in u, let the default setting do the job.

As for DSCP based WRED, it drops excess packets in a congested link, starting with lower priority data according to dscp values (default).

also have in mind it happenes in a Per Class basis. that means for exmple, if Best_Effort class uses all its 19% allocated bw and there is no more bandwidth to be used, it starts to drop selectively between af11, 12, 13. I believe it doesn't compare packets between two different classes (BULK / BUSINESS). So to give better service to Business, that's a matter of bandwidth allocation. you can also police the lower priority data classes so they cant use more than what you want them to use, leaving more BW for other important classes.

you could use : random-detect dscp-based in the classes inc. packets w/ different dscp packets.

Hope it Helps,

Soroush.

[ctranminh]

Thanks for your reply. I now understand!

I see what you mean, each class has its own WRED profile. I also reread my book and highlight that statement .

In  the past I saw all packets at last go into the hardware queue, then I  want to clearly differentiate each type of packet. However I forgot  before that I have already allocated each class with a private BW.

[ctranminh]

One more question about CB-WRED

The current average queue size use to determine the drop probability, not the current queue size so the Average here is Sum of number type-af11 packets, number of type-af12 packets and type-af13 packets divide to 3?

It also mean one class (for example BULK_DATA) have only one queue (defaut queue limit = 64 packets) and all kind of af11 12 13 go into the queue when congesting?

Do I go right? Plz help me

[smehrnia]

Hi,

Average queue size, is a dynamic value that the devices calculates each time a packet becomes candidate for Interface's output queue. how much this value gets close to Max threshold defines the probability of a packet drop in the queue.

in DSCP-WRED you have the ability to define different RED thresholds (min, max) for different dscp's, but i believe if u dont do so, all dscp in one class is treated as if they're the same. i mean it wouldnt essentialy devide the queue between 3, i.e. in the Best_Effort class.

well, the default queue size is 64K, but the queue size is tunable and under certain circumstances this size could be different and become as big as the whole buffer size, causing no buffer drops.

it some how depends on the drop strategies you are using in a class and the underlaying physical interface that u r applying the policy, buffer limis, etc. thats why i told u u have to have through knowledge abt it

Hope it Helps,

Soroush.

[ctranminh]

Thank you very much about your replies..

I am still ambiguous about "Average queue size". but  You are correct, I had better to read more. Actually I have just touch  QoS field for 2 month, and finished one QoS book (student guide).

At summary, by soroushm 's second post I claim the defaut queue  size is 64 packet, and I understand which certain circumstances this  size be different. However I am not sure about each class has only one  queue and all of af11 af12 af13, for example go into that queue when  congestion occur..

Just feel free to answer my question..

Another issue not related to CBWRED, but I post here to easily sum my issues

To  my work, I just put some dscp setting as its default, then do more with  Shaping. Because I am doing QoS between 2 drilling rigs, which use  wireless as the media. The distance is about 7~8 miles. In fact, the congestion not occur in Cisco equipment,  in stead occuring Motorola wireless equipment. My edge router connect to  Motorola tool by fastethernet interface. I need to do a shaping  bandwidth at output interface of Edge router to fit bandwidth of Motorola wireless.  But thing is not easy at all, the configured bandwidth of Motorola  wireless is 10 Mb, for example; however it may change dynimically and  also depend on the wheather. Therefore should I configure shape in average or peak and in percent or explicit bandwidth in bit?