load balancing questions

  • 1
  • Question
  • Updated 4 years ago
  • Answered
Just wondering about the load balancing feature and how it works.  The 'High Density design Guide' seem to imply that load balancing only happens for clients as they are trying to associate, is that correct?

What is an ap has many clients, but not particularly busy, but the suddenly becomes very busy with traffic.  Will those currently connected clients be pushed to another ap or radio?

Is 802.11k important for load balancing?
Photo of wombat

wombat

  • 62 Posts
  • 3 Reply Likes

Posted 4 years ago

  • 1
Photo of Brian Powers

Brian Powers, Champ

  • 396 Posts
  • 92 Reply Likes
There's not a lot of detail as to how Aerohive does load balancing.  If you choose to do it based upon station count, does it do it on a purely 50/50 basis?  As there is no way that I have found to see these statistics. 

There's also not a lot of detail about how it load balances when based off of "Airtime-based". 

Here is a link that details out how another vendor does their aggressive load balancing.  http://www.cisco.com/c/en/us/support/docs/wireless/4400-series-wireless-lan-controllers/107457-load-...

As for 802.11k, since you can enable/disable 802.11k and enable/disable load balancing completely independent of each other, I highly doubt that one relies on the other.  802.11k is an official standard as well whereas load balancing is one of those things that is left up to the vendors to implement how they see fit (I think at least...) similar to how beam forming was not standardized in the 802.11n standard.

Photo of Brian Ambler

Brian Ambler

  • 245 Posts
  • 126 Reply Likes
Hi Michael,

I would have a look first in the help system to see if the information contents the answer you seek, though I can confirm that clients will not be forcibly removed from an AP after they have associated and authenticated for load balancing purposes.

Here is an extract from the help system on load balancing, hopefully it helps you find your answer.
Load Balancing Clients

When an AP radio becomes overloaded with too much wireless traffic, it can use load balancing to manage association requests and clients. Multiple hive members within radio range of one another can distribute clients among themselves. By sharing information about client loads, client airtime consumption, CRC error rates, and RF interference conditions with one another, they build a shared database of AP and network conditions. Based on these conditions, some APs can decide to ignore probe and association requests from new clients so that they will associate with other APs with more favorable conditions.

Enable client load balancing among neighboring hive members: Select to enable client load balancing. Clear to disable it.

Load balancing mode: You can choose how load balance is implemented. If you choose airtime-based load balancing, APs balance the wireless load according to the amount of airtime each client consumes. If you choose station-number load balancing, APs balance the load strictly according to the number of clients that are connected to them.

Airtime-based: Choose this option to cause the APs to balance the wireless load according to the amount of airtime each client consumes by shifting clients that are using much airtime to APs whose airtime load is relatively light.

Station-number:Choose this option to cause the APs to balance the wireless load strictly according to the number of stations connected to the APs. When load balancing by station number, APs migrate clients so that each AP has as close to the same number of clients attached as possible.

Ignore probe and association requests when any of the following limits have been exceeded: The following are the factors APs use to determine whether to accept or ignore probe and association requests from clients.

  • These options only appear if you select airtime-based load balancing.
Maximum CRC error rate per AP: CRC errors occur when the receiver cannot calculate the checksum of a frame accurately. Consequently, in the case of a unicast frame, the receiver does not reply with an acknowledgement, and so the transmitter must resend the frame (multicast and broadcast frames do not require acknowledgement). CRC errors are indicative of possible frame collisions caused by radio interference. If the CRC error rate exceeds the maximum acceptable level for an AP radio, which is 30% by default, then the AP does not accept any more probe and association requests from new clients. You can keep the default setting or modify the threshold from 1 to 99%.

Maximum RF interference per AP: The AP compares the number of signals it receives from sources other than clients on the same frequency as it is using for receiving and transmitting network traffic. If the amount of interference exceeds the maximum acceptable level, which is 40% by default, it accepts no more probe and association requests from new clients. You keep the default setting or modify the threshold from 1 to 99%.

Minimum average airtime per client: This setting considers the average amount of airtime that a client actually consumes in relation to all the available airtime. If that average amount is below a specified threshold—the default is 4%—then the AP does not accept any more probe and association requests from new clients. You can keep the default setting or modify the threshold from 1 to 5%.

Ignore probe and association requests when a client is associated with another hive member until the anchor period elapses: When an AP accepts a new client, it applies an anchor-and-release mechanism to ensure that the newly associated client does not disassociate from its current access point and roam to another one too soon. In a congested RF environment where all the access points have similarly slow wireless links, a client might repeatedly roam in an attempt to improve its data rate. However, with all the access points in the same state, roaming too often does not offer any advantage while possibly causing link instability for the client. The period of time that a client must remain anchored to an access point is called the anchor period.

Anchor Period: Keep the setting at its default value of 60 seconds, or decrease or increase the anchor period from 10 to 600 seconds (10 minutes).

  • An AP can release clients before the anchor period expires if the client's SNR drops below the SNR threshold or if the AP becomes overloaded.
Selectively respond to probe requests if the SNR (Signal-to-Noise Ratio) is stronger than the weak SNR threshold: Select to suppress probe responses if APs detect that the SNR (signal-to-noise ratio) for a client is weak, as determined by its measurement against the SNR threshold.

Weak SNR Threshold: Either keep the SNR threshold set at 15 dB, its default setting, or raise or lower it from 1 to 100 dB. An SNR of 10 dB or lower is generally considered poor and an SNR of 25 dB or greater is generally considered good.

Enable an AP, when it is in an overloaded state or a client's SNR is weak, to respond to association requests after the safety net period elapses: Although the purpose of the settings in the Minimizing Management Traffic section is to increase airtime for user traffic by reducing management traffic, the overarching goal for APs is to provide wireless network connectivity. Select the check box to enable APs that are overloaded or receive association requests from clients with a weak SNR to deflect clients for only a limited period of time—with the intention that they will then associate with a less busy AP or with a closer one—but eventually respond if clients cannot obtain network connections elsewhere. That period during which APs ignore association requests is referred to as a "safety net period". To disable the safety net option, clear the check box.

Safety Net Period: Either keep the safety net period at its default setting of 60 seconds or shorten or lengthen it from 5 to 300 seconds. A longer safety net period might result in longer delays during the initial WLAN connection but greater chances of associating with an AP that can provide superior data rates. On the other hand, a shorter period can accelerate the initial connection but perhaps at the cost of a less optimal data rate.

Original content can be found here

Hope this helps
(Edited)
Photo of J. Goodnough

J. Goodnough, Champ

  • 266 Posts
  • 32 Reply Likes
If you take a look at an APs log while it's load-balancing, you'll see a message like

<182>kernel: [wifi]: wifi1.2: suppress request from 30:f7:c5:xx:xx:xx, reason load-balance-airtime

As Brian Ambler said, APs won't kick a client off, but they will not respond to association requests if their load is too high. You can also issue a show client-load-balance status command to see what's going on an AP at any given time. If all APs are busy, the best AP will eventually respond after the safety net period, according to the documentation - again, as Brian Ambler posted.

Photo of wombat

wombat

  • 62 Posts
  • 3 Reply Likes
That's great info.  Thanks.
Photo of Brian Powers

Brian Powers, Champ

  • 396 Posts
  • 92 Reply Likes
What classifies as "load is to high"?  Wouldn't it be better if that was a configurable option? 

If the same radio profile that has load balancing enabled is assigned to 6 APs that are all within earshot of each other, will clients (regardless of their location in relation to those 6 APs) always be balance between them based on what configurable options there are?  Or is there something else going on that says it'll only balance to a neighbor AP if that AP is within a certain RSSI value of another AP?  Or does it not matter at all and the AP just ignores probes from clients until it receives shared client statistics from neighbors that say they are on even keel again and it should accept the next association request it hears...

There are just to many unknowns for me...
Photo of J. Goodnough

J. Goodnough, Champ

  • 266 Posts
  • 32 Reply Likes
What classifies as "load is to high"?

from the documentation

Maximum CRC error rate per AP: CRC errors occur when the receiver cannot calculate the checksum of a frame accurately. Consequently, in the case of a unicast frame, the receiver does not reply with an acknowledgement, and so the transmitter must resend the frame (multicast and broadcast frames do not require acknowledgement). CRC errors are indicative of possible frame collisions caused by radio interference. If the CRC error rate exceeds the maximum acceptable level for an AP radio, which is 30% by default, then the AP does not accept any more probe and association requests from new clients. You can keep the default setting or modify the threshold from 1 to 99%.
Maximum RF interference per AP: The AP compares the number of signals it receives from sources other than clients on the same frequency as it is using for receiving and transmitting network traffic. If the amount of interference exceeds the maximum acceptable level, which is 40% by default, it accepts no more probe and association requests from new clients. You keep the default setting or modify the threshold from 1 to 99%.
Minimum average airtime per client: This setting considers the average amount of airtime that a client actually consumes in relation to all the available airtime. If that average amount is below a specified threshold—the default is 4%—then the AP does not accept any more probe and association requests from new clients. You can keep the default setting or modify the threshold from 1 to 5%.
Ignore probe and association requests when a client is associated with another hive member until the anchor period elapses: When an AP accepts a new client, it applies an anchor-and-release mechanism to ensure that the newly associated client does not disassociate from its current access point and roam to another one too soon. In a congested RF environment where all the access points have similarly slow wireless links, a client might repeatedly roam in an attempt to improve its data rate. However, with all the access points in the same state, roaming too often does not offer any advantage while possibly causing link instability for the client. The period of time that a client must remain anchored to an access point is called the anchor period.

Suppose all APs in a hive are configured with airtime balancing. A client will attempt to associate to whatever AP has the strongest signal; this AP may suppress the request due to its load-balancing settings. This will set the safety-net timer. If the safety-net timer is exceeded and the AP receives an association request, it will accept the association even though it is in violation of its airtime balancing settings. This is also configurable.

It doesn't particularly matter if all APs in a hive are configured for load-balancing; a client can only attempt to associate to whatever APs it can see regardless of load-balancing settings.

To be fair, I haven't tested this in detail with sniffing and all, but this is based on my understanding of the documentation and inspection of AP logs.
Photo of Brian Powers

Brian Powers, Champ

  • 396 Posts
  • 92 Reply Likes
Those are for "Airtime-Based" load balancing.  What about station number? 

All it seems to state is:

Anchor Period: Keep the setting at its default value of 60 seconds, or decrease or increase the anchor period from 10 to 600 seconds (10 minutes).

  • An AP can release clients before the anchor period expires if the client's SNR drops below the SNR threshold or if the AP becomes overloaded.


Again it states details about an AP becoming "overloaded", but never is that term defined.  Is 20 clients overloaded...  30?  50?  Does it base it per radio or per AP?  Based on channel utilization, airtime statistics, etc.?


And in full disclosure, I've done little testing with the load balancing myself either. 
Photo of Brian Ambler

Brian Ambler

  • 245 Posts
  • 126 Reply Likes
You know, that is a very good question! I don't actually know myself as it was not something that frequently came up with customers and is very difficult to lab up due to the number of clients that would need to be involved. I am relatively certain that there is some threshold, as (from your previous example) the AP never seemed to load balance with six connected clients, but I honestly do not know what that limit is, if it does exist.

Let me reach out to a few colleagues around the office and see what I can come up with as far as specifics.
Photo of J. Goodnough

J. Goodnough, Champ

  • 266 Posts
  • 32 Reply Likes
I suspect the definition of "overloaded" would match the conditions in which the AP would go into association suppression in the first place, which would mean per radio.

I would also agree that hive-wide station-number-balance is probably a bad idea, though station-number-balance is pretty much inferior to airtime-balance anyway.