Port Load sharing

  • 2 December 2014
  • 44 replies
  • 3408 views


Show first post

44 replies

Userlevel 3
Load sharing is not the same as load balancing. EXOS does load sharing (link aggregation), but not load balancing. As Parthiban and Jarek have mentioned, you can adjust the hash algorithm to attempt to better spread the traffic across the links.

I'd like to try to understand your use case. Is there a problem with the traffic favoring one link over another, or is it just a preference to have it balanced? What type of traffic is between these two switches? Using the L3_L4 algorithm, you could see one link more saturated than the other if the IP address and protocol port is the same for a majority of the traffic. If you use the L2 algorithm, it will be based on the MAC address of the source and destination systems - if they are the same, only one link will be chosen.
Hi Drew,

Could you please explain more about upgrade the links to 10G via XGM module for X450a?

current model is 24-port 1000BASE-X SFP plus 4-port 10/100/1000BASE-T.
Anyway can we upgrade to 10G uplink?

BRdgs,
Paul
Userlevel 7
Load sharing is not the same as load balancing. EXOS does load sharing (link aggregation), but not load balancing. As Parthiban and Jarek have mentioned, you can adjust the hash algorithm to attempt to better spread the traffic across the links.

I'd like to try to understand your use case. Is there a problem with the traffic favoring one link over another, or is it just a preference to have it balanced? What type of traffic is between these two switches? Using the L3_L4 algorithm, you could see one link more saturated than the other if the IP address and protocol port is the same for a majority of the traffic. If you use the L2 algorithm, it will be based on the MAC address of the source and destination systems - if they are the same, only one link will be chosen.
Hi Paul,
You'll need to purchase a module and optics for installing into the slot on the back of each switch. For the X450a, I would get the XGM2-2sf. It looks like this:


Once installed, it's a simple configuration to move over to the new 10G link(s).
Userlevel 3
Load sharing is not the same as load balancing. EXOS does load sharing (link aggregation), but not load balancing. As Parthiban and Jarek have mentioned, you can adjust the hash algorithm to attempt to better spread the traffic across the links.

I'd like to try to understand your use case. Is there a problem with the traffic favoring one link over another, or is it just a preference to have it balanced? What type of traffic is between these two switches? Using the L3_L4 algorithm, you could see one link more saturated than the other if the IP address and protocol port is the same for a majority of the traffic. If you use the L2 algorithm, it will be based on the MAC address of the source and destination systems - if they are the same, only one link will be chosen.
Thanks Drew,

It's very informative and a good solution for users.
Is it 2x 10G port ?
Since we are now in this stage.
Please kindly share the step on simple configuration.
We need the down time for this module to installation right.
Userlevel 7
Load sharing is not the same as load balancing. EXOS does load sharing (link aggregation), but not load balancing. As Parthiban and Jarek have mentioned, you can adjust the hash algorithm to attempt to better spread the traffic across the links.

I'd like to try to understand your use case. Is there a problem with the traffic favoring one link over another, or is it just a preference to have it balanced? What type of traffic is between these two switches? Using the L3_L4 algorithm, you could see one link more saturated than the other if the IP address and protocol port is the same for a majority of the traffic. If you use the L2 algorithm, it will be based on the MAC address of the source and destination systems - if they are the same, only one link will be chosen.
Hi Paul,
Yes - that module is 2x 10G SFP+ ports. You will need to power off each switch to install it.

The configuration is easy to do, but will require looking at the existing VLAN config for the current LAG ports. I feel like that is outside the scope of the community here and would be best left to GTAC to help. If you're already planning on purchasing the modules and optics (be sure to get Extreme Networks Certified Optics!), go ahead and open a new case with GTAC to request help moving the configuration from the two LAG ports to the two new ones - that's the best way to prepare.

-Drew
Userlevel 2
To add to Drews comment: LACPs main use case is to add/improve redundancy. That it does some kind of load sharing is a nice benefit, not more. If real load balancing is necessary some kind of L3 routing protocol, for example OSPF, would be a solution. Some routing protocols (if not all; I am no routing expert) offer "equal path load balancing" which actually looks at the link usage and actively distributes the load.
Userlevel 4
As far as I know .Multicast traffic will not be load shared in a lag .
It will always take one link in a lag.There were known PD's
may be this feature should have been implemented in new exos release but not till 15.3 I guess
Userlevel 7
As far as I know .Multicast traffic will not be load shared in a lag .
It will always take one link in a lag.There were known PD's
may be this feature should have been implemented in new exos release but not till 15.3 I guess
This statement is misleading. Multicast can be hashed and load shared in a lag, but because much of the traffic is the same source and destination, a single stream can "stick" to one link. Different streams should hash to different links, unless some aspect of the traffic puts them together due to hashing.
Userlevel 4
As far as I know .Multicast traffic will not be load shared in a lag .
It will always take one link in a lag.There were known PD's
may be this feature should have been implemented in new exos release but not till 15.3 I guess
I am 100% sure there was a PD which said multicast traffic will not load share in a lag.

In 15.6 Concepts guide it is documented that this is supported

Link Aggregation AlgorithmsSummitStack supports address-based load sharing. (This platform does not support port-based load
sharing.)
The following are the types of traffic to which addressed-based algorithms apply and the traffic
components used to select egress links:
• Layer 2 frames and non-IP traffic—The source and destination MAC addresses.
• IPv4 and IPv6 packets
• L2 algorithm—Layer 2 source and destination MAC addresses. Available on SummitStack and all
Summit family switches.
• Broadcast, multicast, and unknown unicast packets (not configurable)—Depends on traffic type:
• IPv4 and IPv6 packets—The source and destination IP addresses.
• Non-IP traffic—The source and destination MAC addresses.
You can control the field examined by the switch for address-based load sharing when the load-sharing
group is created by using the following command:
Userlevel 7
As far as I know .Multicast traffic will not be load shared in a lag .
It will always take one link in a lag.There were known PD's
may be this feature should have been implemented in new exos release but not till 15.3 I guess
I can't find it at the moment, but that doesn't mean it doesn't exist...

Since this is an X450a, the last supported version is 15.3.x. Paul, I would suggest first making sure you're on the most recent software for your cores. At the moment, that is EXOS v15.3.4.6-patch1-8. If there's a software bug, it will hopefully be resolved by updating. TAC will need to manage the case beyond that.
Userlevel 3
As far as I know .Multicast traffic will not be load shared in a lag .
It will always take one link in a lag.There were known PD's
may be this feature should have been implemented in new exos release but not till 15.3 I guess
Hi Drew,

We are not able to purchase a module XGM2-2sf for the X450a. They said it's EOL. Any idea for this where to order please.

Thanks.
Userlevel 7
As far as I know .Multicast traffic will not be load shared in a lag .
It will always take one link in a lag.There were known PD's
may be this feature should have been implemented in new exos release but not till 15.3 I guess
Hi Paul,
There are other outside outlets for purchasing older gear that you can find by doing some Googling. You should be able to find an option in your region that carries this part and the optics. If not, your local Extreme Networks sales team may have some other recommendations for upgrade. Sorry I can't be of further help on this question right now.
Hello guys,

i can share some experience regarding this case. I have made a sharing group with 6x10G ports between a stack topology with 4 nodes and a X670 switch. The problem was that the first and the second ports were going at peak time up to flat 10G and the other ports at that time were loaded at 5G for example. My XOS version is 15.3.2.11.
In my case the solution was :
1. change the custom address-based algorithm from XOR to CRC-16
2. disable the sharing group and creating new one with address-algorithm custom

Output from the device:

configure sharing address-based custom hash-algorithm crc-16
disable sharing 30
enable sharing 30 grouping 30-35 algorithm address-based custom lacp

Hope this was useful !

Best regards,

Venko Velev
@ Venko Velev

So I assume that after you did the mentioned changes you were able to literally do a load-balance on the links in the LAG group.

Regards,
Manish
Userlevel 1
Hi Guys,
This whole thread is quite informative and helped me to correct my understanding on Load sharing on Extreme x450 switches.

But i have couple of questions.

Brief:
In my case, two x450a connected to each other; a static load-sharing (2 physical ports, address based L2) is running and only one of the port is 80% utilized while the 2nd port in LAG has less than 5%. I can understand (or better word is assume) that majority of the traffic has same source and destination MAC, though I do not have wireshark or other logs to support this statement.

Problematic scenario:
There were massive outages reported and during troubleshooting on these Extreme switches, it was found that utilization of one port in load sharing touched 100% but for the 2nd port, it remained same as earlier.

Few of activities performed then:
1) disable/enable of 2nd port on both switches but no change.
2) Disable/enable sharing on one switch
2) added 3rd link in LAG which started carrying the traffic but the older 2nd port of LAG still exhibited same behavior.
-------
We then received an update that the issue is resolved. I am now not sure if 2nd or 3rd activity or some activity at any of the other nodes (Not Extreme switches but Application servers- yes there were few activities carried out on Application servers by other teams during the same time) solved the issue.
------

3) We then changed the configuration of LAG from L2 to L3_L4. Traffic pattern and nodes are under observation now.

Now comes the real question for two configurations ((sharing L2 as well L3_L4).

1) What will the behavior in case utilization of one of the ports in Load sharing touch 100%?
2) Will the extra traffic be dropped or shifted to the 2nd port?
3) How to check if the over utilized port in LAG is dropping traffic?

Your inputs are much awaited on this case.

Thanks
Harkanwal
Userlevel 7
Traffic would be dropped. Changing the hashing algo to l3_l4 is a good try. If most of your traffic is between 2 end-systems, you need to find some entropy in the headers to help load-balancing. L4 usually helps there.
Userlevel 6
In addition to comments from Stephane, the packets at the port level can be monitored using the command, "show port congestion".
Router---------Extreme switch(X450a-48t)--------(Port aggregation)-------Extreme Switch (Summit48si),

On X450a-48t (enable sharing 2:2 grouping 2:2-4 algorithm address-based L3_L4) and on Summit48si (enable sharing 1 grouping 1,2,3 algorithm address-based )

We are facing congestion on all aggregated ports (X450a-48t) while not able to see on Summit48si. There is sufficient bandwidth. What could be the problem
Userlevel 3
Hi,

congestion counter increase when switch buffer on egress have insufficient space for packets.

Mostly traffic flows from router to access switch and because of that you see congestion only on X450a.
For example you have about 500 Mb egress traffic per 1G port.
Internet traffic is not constant, some time we see micoroburst - we try send more than 1G or near 1G

What you see after: 'show port buffer' ?

What you can do:
- enable flow control
- if you use qosprofiles - delete unused qosprofiles
- tune port buffer - don't change if you don't know what are you doing :)

--
Jarek

Reply