Hi Felix,
I've tried testing both AA r38347 as-is, and also AA recompiled with the
back-ported copies of hostapd and mac80211 packages that you provided on
your git repo (thank you!). Unfortunately, in both instances, I saw the
adhoc link between two Engenius EOC-1650 freeze entirely during a long wget
transfer. That is, the transfer itself stalls, even ping packages don't
pass. Several minutes after stopping the wget transfer, the adhoc returns
to normal.
Here are rate control stats on the remote node (i.e. the one not connected
to a wired LAN) before and after attempting the long wget transfer:
BEFORE
root@WasabiNet-mushmaus:~# cat
/sys/kernel/debug/ieee80211/phy0/netdev\:wlan0-1/stations/06\:02\:6f\:76\:13\:53/rc_stats
rate throughput ewma prob this prob this succ/attempt success
attempts
1 0.0 0.0 0.0 0( 5)
14 5783
2 0.7 37.3 0.0 0( 1)
548 956
P 5.5 4.9 91.3 100.0 1( 1)
158 395
D 11 6.0 59.1 100.0 0( 0)
267 643
6 1.8 30.6 0.0 0( 0)
231 508
9 3.3 37.2 0.0 0( 0)
349 727
12 4.6 39.3 33.3 0( 0)
641 1106
C 18 6.0 34.8 0.0 0( 0)
351 698
B 24 10.6 46.9 100.0 1( 1)
63 326
36 4.6 14.2 0.0 0( 6)
156 627
48 0.0 0.1 0.0 0( 0)
771 2441
A 54 24.2 52.3 33.3 1( 3)
842 2721
Total packet count:: ideal 5951 lookaround 664
AFTER
root@WasabiNet-mushmaus:~# cat
/sys/kernel/debug/ieee80211/phy0/netdev\:wlan0-1/stations/06\:02\:6f\:76\:13\:53/rc_stats
rate throughput ewma prob this prob this succ/attempt success
attempts
1 0.0 0.0 0.0 0( 6)
18 25581
2 0.3 18.2 0.0 0( 0)
564 1122
5.5 3.3 62.0 100.0 0( 0)
476 1159
11 5.5 54.4 100.0 0( 0)
339 1024
6 2.2 37.7 100.0 0( 0)
3900 5892
9 3.1 35.3 0.0 0( 0)
614 1313
D 12 7.9 67.0 100.0 0( 0)
660 1374
C 18 9.6 55.8 50.0 0( 0)
536 1509
24 3.7 16.5 33.3 0( 0)
264 1248
36 6.7 20.9 0.0 0( 0)
166 867
A P 48 27.9 67.2 100.0 0( 0)
1682 5160
B 54 13.4 29.0 33.3 0( 0)
3788 10223
Total packet count:: ideal 8510 lookaround 946
On both devices running AA r36669, long wget transfers work fine. In my
instance, the transfers averaged to 415Bytes/sec over a single test that
moved 2GBytes. For comparison, below are the rc_stats on this same device
running AA r36669, before and after a successful long transfer. I do notice
that r36669 reports 0 throughput at higher rates like 54Mbit/s, unlike
r38347. Maybe throughput is being measured inaccurately?
BEFORE
root@WasabiNet-mushmaus:~# cat
/sys/kernel/debug/ieee80211/phy0/netdev\:wlan0-1/stations/06\:02\:6f\:76\:13\:53/rc_stats
rate throughput ewma prob this prob this succ/attempt success
attempts
1 0.5 52.2 100.0 0( 0)
25 39
2 1.2 63.0 100.0 0( 0)
4 7
5.5 3.1 58.9 100.0 0( 0)
4 6
B P 11 8.7 86.0 100.0 0( 0)
10 11
6 4.0 66.8 100.0 0( 0)
5 6
9 4.7 52.5 0.0 0( 0)
79 167
C 12 7.3 62.2 100.0 0( 0)
4 11
A 18 11.7 67.5 50.0 1( 2)
587 969
D 24 5.0 22.2 19.9 0( 0)
4 34
36 0.0 0.0 0.0 0( 0)
0 12
48 0.0 0.0 0.0 0( 0)
0 12
54 0.0 0.0 0.0 0( 0)
0 12
Total packet count:: ideal 653 lookaround 72
AFTER
root@WasabiNet-mushmaus:~# cat
/sys/kernel/debug/ieee80211/phy0/netdev\:wlan0-1/stations/06\:02\:6f\:76\:13\:53/rc_stats
rate throughput ewma prob this prob this succ/attempt success
attempts
1 0.0 0.0 0.0 0( 0) 1657
32236
2 1.2 60.5 100.0 0( 0) 8632
16530
5.5 1.6 31.3 0.0 0( 0) 14387
31865
C 11 4.9 48.4 8.3 0( 0) 107143
238102
6 2.9 49.4 100.0 1( 1) 5839
20081
B 9 5.2 58.0 50.0 0( 0) 10665
38516
12 3.3 28.1 0.0 0( 0) 40247
120171
A P 18 12.6 73.0 72.2 13( 18) 369301
883412
D 24 4.6 20.3 33.3 0( 0) 296475
1216256
36 0.0 8.3 0.0 0( 0) 37298
330903
48 0.0 9.3 0.0 0( 0) 13596
222798
54 0.0 7.7 0.0 0( 0) 34449
388569
Total packet count:: ideal 8629 lookaround 958
On Sun, Oct 13, 2013 at 1:21 PM, Felix Fietkau <[email protected]> wrote:
> On 2013-10-13 7:49 PM, Ben West wrote:
> > The devices in 'production' use are Engenius EOC-01650 and Open Mesh
> > OM1Ps, both with Atheros SoC AR2315. These are gradually by being
> > replaced by UBNT Nanostation Loco M2's, with SoC AR7240.
> >
> > The small adhoc network I was using for proving firmware, where the
> > decrease in throughput was observed, is comprised of 3 EOC-1650's, hung
> > up at various locations around my building.
> >
> > My simple speed test was just to wget a 500Mbyte file from a 100Mbit
> > wired LAN connected to one of the nodes across the adhoc network to
> > /dev/null. Indeed, iperf would be more precise, but wget seemed
> > sufficient just to allow me to observe large differences in throughput,
> > e.g. 1Mbit/s vs 3Mbit/s average.
> >
> > At any rate, is it preferred to compare throughput values I've observed
> > using AA r36669 with those observed running AA r38346, or with
> > throughput values measured using current trunk. I understand that since
> > AA only receives a selection of backports from trunk, it can
> > occasionally be a hodgepodge of working vs suboptimal code.
> You can use the full trunk backport by using the package from this
> repository: http://nbd.name/gitweb.cgi?p=aa-mac80211.git;a=summary
> I'd recommend comparing that with the older version.
> In addition to the different throughput values, please also provide rate
> control statistics from
> /sys/kernel/debug/ieee80211/phy0/netdev:wlan0/stations/*/rc_stats
>
> Thanks,
>
> - Felix
>
--
Ben West
http://gowasabi.net
[email protected]
314-246-9434
_______________________________________________
openwrt-devel mailing list
[email protected]
https://lists.openwrt.org/cgi-bin/mailman/listinfo/openwrt-devel
