Hi Yeoh,

That's great info, thanks!  We'll ask our ath9k experts if they can
help us enable hw acceleration in mesh mode.

Cheers,

Javier

On Mon, Nov 14, 2011 at 11:42 PM, Yeoh Chun-Yeow <[email protected]> wrote:
> Hi, Javier
>
> I have tried to setup one AP and one STA and do the throughput measurements
> and compare with Mesh setup.
>
> Based on the results, I observed that AP/STA mode using PSK2 has suffered
> the same problem by loading the ath9k kernel module with nohwcrypt=1. It
> registered the throughtput of ~14Mbps (HT20) if the nohwcrypt=1. For
> nohwcrypt=0, it can achieve up to 50Mbps. There is additional process for tx
> and rx, such as:
>
> .....
>         pos += CCMP_HDR_LEN;
>         ccmp_special_blocks(skb, pn, scratch, 0);
>         ieee80211_aes_ccm_encrypt(key->u.ccmp.tfm, scratch, pos, len,
>                                   pos, skb_put(skb, CCMP_MIC_LEN));
> .....
>
> So software ccmp encryption and decryption is perhaps the bottleneck.
>
> Regards,
> Chun-Yeow
>
> AP/STA Mode: (nohwcrypt=1)
> # iperf -c 10.44.29.1 -i 1 -t 10
> [  3]  0.0-10.1 sec  17.9 MBytes  14.8 Mbits/sec
>
> AP/STA Mode: (nohwcrypt=0)
> # iperf -c 10.44.29.1 -i 1 -t 10
> [  3]  0.0-10.0 sec  50.5 MBytes  42.2 Mbits/sec
>
> None-Secured Mesh:
> # iperf -c 10.44.29.1 -i 1 -t 10
> [  3]  0.0-10.1 sec  34.6 MBytes  28.9 Mbits/sec
>
> Secured Mesh: (nohwcrypt=1)
> # iperf -c 10.44.29.1 -i 1 -t 10
> [  3]  0.0-10.1 sec  17.3 MBytes  14.3 Mbits/sec
>
>
>
> On Tue, Nov 15, 2011 at 5:40 AM, Javier Cardona <[email protected]> wrote:
>>
>> Hi Yeoh,
>>
>> The authentication daemon is not in the data path:  its role ends once
>> a new candidate has been authenticated, keys derived and peering
>> established.  These are the things we investigated, without much
>> success:
>>
>> Q1: Are HT data rates used for secured mesh traffic?
>> A1: Yes, we confirmed that.
>>
>> Q2: Are there more retries/corrupted frames on secured traffic vs.
>> open mesh traffic?
>> A2:  No.  In our environment we observed roughly 10% retransmission rates.
>>
>> Q3. Is encryption/decryption bound by CPU?
>> A3: We did observe a 10% higher CPU usage when running mesh in secure
>> mode.  But our host CPU had ample available cycles to handle that.
>>
>> Q4: Is software encryption/decryption adding latency to the datapath
>> and hence decreasing throughput?
>> A4: We did not observe any additional inter-frame latency.
>>
>> Q5. Is this HT-related?
>> A4: No, we have also observed decreased throughput in non-HT modes.
>>
>> To summarize:  We have also observed the decreased throughput in
>> secured mode (in our setup, ~50% throughput reduction) and so far we
>> have not found where the bottleneck is yet.
>>
>> Cheers,
>>
>> Javier
>>
>> On Sun, Nov 13, 2011 at 6:05 PM, Yeoh Chun-Yeow <[email protected]>
>> wrote:
>> > Hi, Thomas
>> >
>> > Managed to get the secured mesh up after applying the linked patch. I
>> > have
>> > managed to test the HT20 mode for secured mesh and non-secured mesh due
>> > to
>> > my embedded platform only supports Fast Ethernet. Here are the results:
>> >
>> > Non Secured Mesh (HT20)
>> > # iperf -s -i 1
>> > ------------------------------------------------------------
>> > Server listening on TCP port 5001
>> > TCP window size: 85.3 KByte (default)
>> > ------------------------------------------------------------
>> > [ ID] Interval       Transfer     Bandwidth
>> > [SUM]  0.0-10.0 sec  72.0 MBytes  60.1 Mbits/sec
>> >
>> > Secured Mesh (HT20)
>> > # iperf -s -i 1
>> > ------------------------------------------------------------
>> > Server listening on TCP port 5001
>> > TCP window size: 85.3 KByte (default)
>> > ------------------------------------------------------------
>> > [ ID] Interval       Transfer     Bandwidth
>> > [SUM]  0.0-10.1 sec  15.6 MBytes  13.0 Mbits/sec
>> >
>> > A bit disappointed that the throughput is drop quite significantly.
>> > Confirm
>> > that the aggregation is working:
>> >
>> > Rx A-MPDU request on tid 0 result 0
>> > Open BA session requested for 0a:0b:6b:7d:e3:b9 tid 0
>> > activated addBA response timer on tid 0
>> > switched off addBA timer for tid 0
>> > Aggregation is on for tid 0
>> >
>> > Regards,
>> > Chun-Yeow
>> >
>> > On Wed, Nov 9, 2011 at 5:13 AM, Thomas Pedersen <[email protected]>
>> > wrote:
>> >>
>> >> Hi Chun-Yeow,
>> >>
>> >> Have you had any success?
>> >>
>> >> I did not have time to properly to properly test throughput in a
>> >> secure mesh laste time, but now with our ft-ht-mesh + ft-mesh-fixes
>> >> branches, I'm seeing 55.5Mb/s TCP throughput with a 243Mb/s datarate.
>> >> Throughput in the non-secure case is more like at least 100Mb/s, so it
>> >> seems we have quite a bit of overhead.
>> >>
>> >> Thomas
>> >>
>> >> On Thu, Nov 3, 2011 at 6:28 PM, Yeoh Chun-Yeow <[email protected]>
>> >> wrote:
>> >> > Hi, all
>> >> > Ok. I need to reconfigure authsae in htmode to HT20 and transmission
>> >> > rate
>> >> > increased up to 104Mbps with TCP throughput about 7Mbps.
>> >> > But as Thomas pointed out, aggregation is not working.
>> >> > Regards,
>> >> > Chun-Yeow
>> >> >
>> >> > On Fri, Nov 4, 2011 at 9:04 AM, Yeoh Chun-Yeow
>> >> > <[email protected]>
>> >> > wrote:
>> >> >>
>> >> >> Hi, all
>> >> >> I have captured the packet via Wireshark. It shows that the bit rate
>> >> >> used
>> >> >> by the two mesh nodes is 6Mbps. That's why the throughput is low.
>> >> >> Regards,
>> >> >> Chun-Yeow
>> >> >> On Fri, Nov 4, 2011 at 1:50 AM, Thomas Pedersen <[email protected]>
>> >> >> wrote:
>> >> >>>
>> >> >>> It seems ADDBA requests are dropped by a peer when security is
>> >> >>> enabled, thanks for reporting this!
>> >> >>>
>> >> >>> Thomas
>> >> >>>
>> >> >>> On Thu, Nov 3, 2011 at 9:41 AM, Thomas Pedersen
>> >> >>> <[email protected]>
>> >> >>> wrote:
>> >> >>> > Hi Chun-Yeow,
>> >> >>> >
>> >> >>> >> Using authsae, I only managed to get 4Mbps. However, without
>> >> >>> >> SAE, I
>> >> >>> >> managed
>> >> >>> >> to obtain 60Mbps using iperf 10 parallel TCP streams.
>> >> >>> >> Others are getting the same results? It seems that performance
>> >> >>> >> drop
>> >> >>> >> is
>> >> >>> >> too
>> >> >>> >> significant.
>> >> >>> >
>> >> >>> > Did you set the correct channel type with iw before bringing up
>> >> >>> > authsae?
>> >> >>> >
>> >> >>> > Thomas
>> >> >>> >
>> >> >>
>> >> >
>> >> >
>> >
>> >
>>
>>
>>
>> --
>> Javier Cardona
>> cozybit Inc.
>> http://www.cozybit.com
>
>



-- 
Javier Cardona
cozybit Inc.
http://www.cozybit.com
_______________________________________________
Devel mailing list
[email protected]
http://open80211s.com/mailman/listinfo/devel

Reply via email to