Hello,
We study the effects of 802.11n frame aggregation on throughput. We
noticed a
strange pattern in the MPDU loss within an aggregated frame. It seems
that the
second half of the MPDUs (those with higher sequence numbers) in an
aggregated frame
are more likely to be lost. Is this a known fact or is there any
explanation for it?
For example if 32 frames are aggregated with sequence numbers 100 to 131.
Frames with sequence numbers 100-115 are more likely to be received
correctly
than 116-131.
Best,
Ali
On 25 October 2014 08:28, Ali Abedi <[email protected]> wrote:
> Hi Adrian,
>
> We have a high end spectrum analyzer. So we are sure there is no background
> interference
> We run our experiments in the 5 GHZ spectrum. The channel conditions can
> still vary due to
> the movement of the people in the vicinity of the experiment setup. We
> select a rate that
> experiences at least 20% error on average. Since if the error is 100% or 0%
> it's not interesting
> for us.
>
> My point is if the channel conditions change the distribution of failed
> packets should be uniform.
> The first and second half of the packets have the same chance to be
> received successfully.
Here's a little story.
My first wifi contract had me spend months trying to figure out why an
AP was losing its mind. It'd get stuck in a "stuck beacon" loop and
only a hard powercycle of /all/ of the access points in an area would
clear it.
It turned out that the PCB design had some non-grounded /
non-populated tracks that just "happened" to form a 2GHz resonator.
Once we grounded those tracks, the APs started behaving themselves.
The company in question spent months with high end spectrum analysis
kit in the lab (where it never happened) and underground (where it did
happen.) It's only after they stuck the spectrum analyser probe
_inside the access point_ right up close to the NIC did they see it.
Here's the spectrum analyser traces. You can see the peak.
http://www.creative.net.au/ath/
So, weirder crap has happened.
Which NICs and which MCS rates are you using?
-adrian
The paper mentioned that this happens when the client is mobile.
But I confirm Adrian's observation . This problem happens even
in stationary environments with dynamic channels (e.g., people moving in
the vicinity
of the AP/Client).
Best,
Ali
On 14-10-30 12:11 PM, Adrian Chadd wrote:
> On 30 October 2014 08:48, Ali Abedi <[email protected]> wrote:
>> Hi Adrian,
>>
>>
>> We observed that this can happen for any rate for some SNR values.
>> If the SNR is strong enough for the given MCS this won't happen.
>> But when the SNR approaches the transition region when
>> error rate starts to increase, this problem will be observed.
>>
>> So this can happen even for MCS0->MCS4 when the client is far from the AP
>> and specially when it's moving.
> Right. That's the missing useful information here. :)
>
> Yes, I'd expect this happens whilst the client is moving. The training
> stuff is all done on the beginning of the packet and channel
> conditions aren't adjusted during packet reception - only upon the
> next received packet.
>
> (FYI - I've seen a similar pattern, but when i stand between the AP /
> STA at > MCS13 and start waving my hands around. Just that slight
> change in channel conditions == the above failure.)
>
> So thanks for reminding us that we should take A-MPDU length into
> account in our rate control code. :)
>
>
>
> -adrian
But for this channel conditions should be changing at the scale of 100
micro secs consistently.
On Sat, Oct 25, 2014 at 12:13 AM, Adrian Chadd <[email protected]> wrote:
> It's not completely unsurprising - the initial channel estimate and
> such is done at the beginning of each packet and stays constant. So if
> there's some varying channel conditions that change that during the
> duration of a packet, the tail end is going to end up having less SNR
> and may end up getting more errors.
>
>
> -adrian
>
> On 24 October 2014 09:04, Ali Abedi <[email protected]> wrote:
>> Hello,
>>
>> We study the effects of 802.11n frame aggregation on throughput. We noticed
>> a
>> strange pattern in the MPDU loss within an aggregated frame. It seems that
>> the
>> second half of the MPDUs (those with higher sequence numbers) in an
>> aggregated frame
>> are more likely to be lost. Is this a known fact or is there any explanation
>> for it?
>>
>> For example if 32 frames are aggregated with sequence numbers 100 to 131.
>> Frames with sequence numbers 100-115 are more likely to be received
>> correctly
>> than 116-131.
>>
>>
>> Best,
>> Ali
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-wireless" in
>> the body of a message to [email protected]
>> More majordomo info at http://vger.kernel.org/majordomo-info.html
> _______________________________________________
> ath9k-devel mailing list
> [email protected]
> https://lists.ath9k.org/mailman/listinfo/ath9k-devel
Hi Adrian,
We observed that this can happen for any rate for some SNR values.
If the SNR is strong enough for the given MCS this won't happen.
But when the SNR approaches the transition region when
error rate starts to increase, this problem will be observed.
So this can happen even for MCS0->MCS4 when the client is far from the AP
and specially when it's moving.
Thanks,
Ali
On 14-10-29 04:34 PM, Adrian Chadd wrote:
> Just finally -
>
> MCS20 -> MCS23 are very sensitive to changing channel anything. See if
> you can find or make some required SNR curves for each MCS rate.
>
> So although it doesn't surprise me to find this is happening, it's
> very cute that someone's gone and done the work of figuring out how to
> improve the rate control algorithm to take it into account. I'm kinda
> thinking about how to do that with FreeBSD right now.
>
> Do you get the same pattern on say, MCS0->MCS4 over a 4ms long aggregate frame?
>
>
> -adrian
>
> On 25 October 2014 13:35, Ali Abedi <[email protected]> wrote:
>> Thank you for sharing the story.
>> Even if I consider interference as a possibility, still I can't justify the
>> higher
>> chance of frame loss in the second half of the aggregate frame.
>>
>> We use
>>
>> PCI-express 3 antenna dual band cards
>> product: AR93xx Wireless Network Adapter
>> and/or
>> Atheors AR5B97 which is a 2.4 GHz dual antenna internal card in a laptop
>>
>> we also tried TP-LINK TL-WDN4200 N900 as the receiver.
>>
>> However we see the same results.
>> we mostly use MCS 20-23, sgi = 0, 20 MHz channels.
>>
>> The loss pattern is something like this
>> (each line is an imaginary aggregated frame and each bit is the fate of the
>> MPDU)
>>
>> 11111111111100011000000000000
>> 11111110001101011010000000000
>> 11111000000000000000000000000
>> 11111111111010100000000000000
>> 11111100101010000000000000000
>>
>> The interesting part is that with the start of the next frame error rate
>> goes down initially
>> then it goes up again in the second portion of the packet.
>>
>> Best,
>> Ali
>>
>>
>>
>>
>>
>> On 25/10/2014 2:30 PM, Adrian Chadd wrote:
>>
>> On 25 October 2014 08:28, Ali Abedi <[email protected]> wrote:
>>
>> Hi Adrian,
>>
>> We have a high end spectrum analyzer. So we are sure there is no background
>> interference
>> We run our experiments in the 5 GHZ spectrum. The channel conditions can
>> still vary due to
>> the movement of the people in the vicinity of the experiment setup. We
>> select a rate that
>> experiences at least 20% error on average. Since if the error is 100% or 0%
>> it's not interesting
>> for us.
>>
>> My point is if the channel conditions change the distribution of failed
>> packets should be uniform.
>> The first and second half of the packets have the same chance to be
>> received successfully.
>>
>> Here's a little story.
>>
>> My first wifi contract had me spend months trying to figure out why an
>> AP was losing its mind. It'd get stuck in a "stuck beacon" loop and
>> only a hard powercycle of /all/ of the access points in an area would
>> clear it.
>>
>> It turned out that the PCB design had some non-grounded /
>> non-populated tracks that just "happened" to form a 2GHz resonator.
>> Once we grounded those tracks, the APs started behaving themselves.
>>
>> The company in question spent months with high end spectrum analysis
>> kit in the lab (where it never happened) and underground (where it did
>> happen.) It's only after they stuck the spectrum analyser probe
>> _inside the access point_ right up close to the NIC did they see it.
>>
>> Here's the spectrum analyser traces. You can see the peak.
>>
>> http://www.creative.net.au/ath/
>>
>> So, weirder crap has happened.
>>
>> Which NICs and which MCS rates are you using?
>>
>>
>>
>> -adrian
>>
>>
Just finally -
MCS20 -> MCS23 are very sensitive to changing channel anything. See if
you can find or make some required SNR curves for each MCS rate.
So although it doesn't surprise me to find this is happening, it's
very cute that someone's gone and done the work of figuring out how to
improve the rate control algorithm to take it into account. I'm kinda
thinking about how to do that with FreeBSD right now.
Do you get the same pattern on say, MCS0->MCS4 over a 4ms long aggregate frame?
-adrian
On 25 October 2014 13:35, Ali Abedi <[email protected]> wrote:
> Thank you for sharing the story.
> Even if I consider interference as a possibility, still I can't justify the
> higher
> chance of frame loss in the second half of the aggregate frame.
>
> We use
>
> PCI-express 3 antenna dual band cards
> product: AR93xx Wireless Network Adapter
> and/or
> Atheors AR5B97 which is a 2.4 GHz dual antenna internal card in a laptop
>
> we also tried TP-LINK TL-WDN4200 N900 as the receiver.
>
> However we see the same results.
> we mostly use MCS 20-23, sgi = 0, 20 MHz channels.
>
> The loss pattern is something like this
> (each line is an imaginary aggregated frame and each bit is the fate of the
> MPDU)
>
> 11111111111100011000000000000
> 11111110001101011010000000000
> 11111000000000000000000000000
> 11111111111010100000000000000
> 11111100101010000000000000000
>
> The interesting part is that with the start of the next frame error rate
> goes down initially
> then it goes up again in the second portion of the packet.
>
> Best,
> Ali
>
>
>
>
>
> On 25/10/2014 2:30 PM, Adrian Chadd wrote:
>
> On 25 October 2014 08:28, Ali Abedi <[email protected]> wrote:
>
> Hi Adrian,
>
> We have a high end spectrum analyzer. So we are sure there is no background
> interference
> We run our experiments in the 5 GHZ spectrum. The channel conditions can
> still vary due to
> the movement of the people in the vicinity of the experiment setup. We
> select a rate that
> experiences at least 20% error on average. Since if the error is 100% or 0%
> it's not interesting
> for us.
>
> My point is if the channel conditions change the distribution of failed
> packets should be uniform.
> The first and second half of the packets have the same chance to be
> received successfully.
>
> Here's a little story.
>
> My first wifi contract had me spend months trying to figure out why an
> AP was losing its mind. It'd get stuck in a "stuck beacon" loop and
> only a hard powercycle of /all/ of the access points in an area would
> clear it.
>
> It turned out that the PCB design had some non-grounded /
> non-populated tracks that just "happened" to form a 2GHz resonator.
> Once we grounded those tracks, the APs started behaving themselves.
>
> The company in question spent months with high end spectrum analysis
> kit in the lab (where it never happened) and underground (where it did
> happen.) It's only after they stuck the spectrum analyser probe
> _inside the access point_ right up close to the NIC did they see it.
>
> Here's the spectrum analyser traces. You can see the peak.
>
> http://www.creative.net.au/ath/
>
> So, weirder crap has happened.
>
> Which NICs and which MCS rates are you using?
>
>
>
> -adrian
>
>
It's not completely unsurprising - the initial channel estimate and
such is done at the beginning of each packet and stays constant. So if
there's some varying channel conditions that change that during the
duration of a packet, the tail end is going to end up having less SNR
and may end up getting more errors.
-adrian
On 24 October 2014 09:04, Ali Abedi <[email protected]> wrote:
> Hello,
>
> We study the effects of 802.11n frame aggregation on throughput. We noticed
> a
> strange pattern in the MPDU loss within an aggregated frame. It seems that
> the
> second half of the MPDUs (those with higher sequence numbers) in an
> aggregated frame
> are more likely to be lost. Is this a known fact or is there any explanation
> for it?
>
> For example if 32 frames are aggregated with sequence numbers 100 to 131.
> Frames with sequence numbers 100-115 are more likely to be received
> correctly
> than 116-131.
>
>
> Best,
> Ali
> --
> To unsubscribe from this list: send the line "unsubscribe linux-wireless" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Adrian,
We have a high end spectrum analyzer. So we are sure there is no
background interference
We run our experiments in the 5 GHZ spectrum. The channel conditions can
still vary due to
the movement of the people in the vicinity of the experiment setup. We
select a rate that
experiences at least 20% error on average. Since if the error is 100% or
0% it's not interesting
for us.
My point is if the channel conditions change the distribution of failed
packets should be uniform.
The first and second half of the packets have the same chance to be
received successfully.
Thanks,
Ali
On 25/10/2014 11:19 AM, Adrian Chadd wrote:
> On 24 October 2014 13:42, Ali Abedi <[email protected]> wrote:
>> We don't use a rate adaptation at this moment (i.e., fixed rate) and the
>> setup
>> is stationary. So we expect to see relatively stable channel conditions.
>> Even if the channel
>> conditions change during the aggregated frame. The first half of the MPDUs
>> have the same chance of experiencing worse channel conditions.
> How do you /know/ you have stable channel conditions?
>
> There are a lot of things that could be going on inside the devices
> you're testing on. It doesn't have to be channel noise coming in an
> antenna. For example, your computer could be generating rapidly
> changing noise spurs from some clocking sources.
>
> Try firing up the spectral scan mode on the NIC and plot the data. See
> if there are any abnormal peaks going on over time.
>
> And a large / long A-MPDU could be measured in milliseconds of length.
> the original poster didn't say which rate(s) they are trying with and
> how much margin (SNR) the receiver is seeing. Pulling out EVM from the
> received A-MPDU frames would also be helpful.
>
> Thanks,
>
>
>
> -adrian
>
>
>
>
> -adrian
On Sat, Oct 25, 2014 at 12:53 AM, Kamran Nishat <[email protected]> wrote:
>
> But for this channel conditions should be changing at the scale of 100
> micro secs consistently.
>
> On Sat, Oct 25, 2014 at 12:13 AM, Adrian Chadd <[email protected]> wrote:
> > It's not completely unsurprising - the initial channel estimate and
> > such is done at the beginning of each packet and stays constant. So if
> > there's some varying channel conditions that change that during the
> > duration of a packet, the tail end is going to end up having less SNR
> > and may end up getting more errors.
> >
> >
> > -adrian
> >
> > On 24 October 2014 09:04, Ali Abedi <[email protected]> wrote:
> >> Hello,
> >>
> >> We study the effects of 802.11n frame aggregation on throughput. We noticed
> >> a
> >> strange pattern in the MPDU loss within an aggregated frame. It seems that
> >> the
> >> second half of the MPDUs (those with higher sequence numbers) in an
> >> aggregated frame
> >> are more likely to be lost. Is this a known fact or is there any explanation
> >> for it?
> >>
> >> For example if 32 frames are aggregated with sequence numbers 100 to 131.
> >> Frames with sequence numbers 100-115 are more likely to be received
> >> correctly
> >> than 116-131.
> >>
> There is no known limitation/explanation for losing 2nd half of MPDU's in an A-MPDU
*every time (most likely)*. This might specific to the case, can you
share a capture?
On 30 October 2014 08:48, Ali Abedi <[email protected]> wrote:
> Hi Adrian,
>
>
> We observed that this can happen for any rate for some SNR values.
> If the SNR is strong enough for the given MCS this won't happen.
> But when the SNR approaches the transition region when
> error rate starts to increase, this problem will be observed.
>
> So this can happen even for MCS0->MCS4 when the client is far from the AP
> and specially when it's moving.
Right. That's the missing useful information here. :)
Yes, I'd expect this happens whilst the client is moving. The training
stuff is all done on the beginning of the packet and channel
conditions aren't adjusted during packet reception - only upon the
next received packet.
(FYI - I've seen a similar pattern, but when i stand between the AP /
STA at > MCS13 and start waving my hands around. Just that slight
change in channel conditions == the above failure.)
So thanks for reminding us that we should take A-MPDU length into
account in our rate control code. :)
-adrian
On 24 October 2014 13:42, Ali Abedi <[email protected]> wrote:
> We don't use a rate adaptation at this moment (i.e., fixed rate) and the
> setup
> is stationary. So we expect to see relatively stable channel conditions.
> Even if the channel
> conditions change during the aggregated frame. The first half of the MPDUs
> have the same chance of experiencing worse channel conditions.
How do you /know/ you have stable channel conditions?
There are a lot of things that could be going on inside the devices
you're testing on. It doesn't have to be channel noise coming in an
antenna. For example, your computer could be generating rapidly
changing noise spurs from some clocking sources.
Try firing up the spectral scan mode on the NIC and plot the data. See
if there are any abnormal peaks going on over time.
And a large / long A-MPDU could be measured in milliseconds of length.
the original poster didn't say which rate(s) they are trying with and
how much margin (SNR) the receiver is seeing. Pulling out EVM from the
received A-MPDU frames would also be helpful.
Thanks,
-adrian
-adrian
We don't use a rate adaptation at this moment (i.e., fixed rate) and the
setup
is stationary. So we expect to see relatively stable channel
conditions. Even if the channel
conditions change during the aggregated frame. The first half of the MPDUs
have the same chance of experiencing worse channel conditions.
Thank,
Ali
On 14-10-24 03:13 PM, Adrian Chadd wrote:
> It's not completely unsurprising - the initial channel estimate and
> such is done at the beginning of each packet and stays constant. So if
> there's some varying channel conditions that change that during the
> duration of a packet, the tail end is going to end up having less SNR
> and may end up getting more errors.
>
>
> -adrian
>
> On 24 October 2014 09:04, Ali Abedi <[email protected]> wrote:
>> Hello,
>>
>> We study the effects of 802.11n frame aggregation on throughput. We noticed
>> a
>> strange pattern in the MPDU loss within an aggregated frame. It seems that
>> the
>> second half of the MPDUs (those with higher sequence numbers) in an
>> aggregated frame
>> are more likely to be lost. Is this a known fact or is there any explanation
>> for it?
>>
>> For example if 32 frames are aggregated with sequence numbers 100 to 131.
>> Frames with sequence numbers 100-115 are more likely to be received
>> correctly
>> than 116-131.
>>
>>
>> Best,
>> Ali
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-wireless" in
>> the body of a message to [email protected]
>> More majordomo info at http://vger.kernel.org/majordomo-info.html