Hi,

I'm testing a setup with an RB912 with a Ubiquiti SR71e Mini-PCI card.
(I'm using the SR71e in my tests, the RB912's on-board radio is disabled).
I have 2 low-gain omni antennas, and a laptop several metres away.

I don't seem to be able to get TCP throughput (measured by scp-ing a big file) greater than about 5.0MB/s (this is the max, it's often less).
When I remove the RB912 and replace with a UniFi access-point (leaving everything else the same) I get roughly 8.5MB/s -- so I know higher speeds are possible.

I've checked the following on my RB912/SR71e setup:
- I have an 802.11n connection
- I've configured:
- SR71e in ap-bridge mode
- wireless protocol 802.11
- 2GHz-only-N
- 20/40MHz-HT Below
- Both chain0 and chain1 for Tx and Rx
- Otherwise everything is default

Is there some other config option I should be checking, or is this about the best I can expect?


Thanks in advance,

Alex

I did some more digging on this.
Quick summary:
- by setting distance to 'indoors' I can improve throughput a bit but still not excellent
- throughput is heavily asymmetric (downlink faster than uplink)

Detail:
I notice that I'm not the only one, there are similar reports:
http://forum.mikrotik.com/viewtopic.php?f=7&t=57942
http://forum.mikrotik.com/viewtopic.php?f=7&t=35767
(I realise that many people had problems because the encryption was not set to WPA2, mine is WPA2).

I tried some suggestions from the topics above:
1) setting the distance to 'indoors' (seems to improve uplink throughput by ~10Mbps)
2) reducing the hw-retries (doesn't seem to make a big difference)
3) there are suggestions that mikrotik performance is poor on a single TCP stream but good on 10+. Running 20 streams seems to improve throughput by ~10Mbps.

Test methodology: I'm using "iperf -P 20 -c <wired_host> -r -t 60"
(20 parallel streams, testing both tx and rx one after the other)
Results (bandwidth order is STA->AP Mbps / AP->STA Mbps):
- UniFi running 802.11n for comparison: 74 / 73
- MT with "distance=dynamic": 36 / 73
- MT with "distance=indoors": 44 / 76
- MT with "distance=indoors", single TCP stream: 37 / 68

So the best I can do is roughly equal to the UniFi on AP->STA but just about half the UniFi's throughput STA->AP.
I really really want to use the MT because the interface and tools are so much better, but not if there's a big performance disadvantage -- does anyone have any other suggestions of things to try?

(I should also mention that the problem isn't CPU-limited, CPUs are all lightly-loaded)

TCP perfomance is very sensitive to packet latency jitter (ping jitter). Bringing the jitter down will improve TCP performance.
What is ping jitter? In case you do "ping host" ping times vary much. Jitter is a real problem on wireless links.

For two 5 GHz Mikrotik-Sxt's using NV2 on a link I was able to bring the jitter down doing the following:
- setting NV2 TDMA period to 1 ms, cell radius to lowest possible value (10 km)
- reduce "HT MCS" so that TX/RX CCQ is near 100 (this will reduce packet retransmits)
HT MCS is directly related to the maximal "TX/RX rate" shown in the registration tab of the wireless table.
E.g. on a 5 GHz Sxt (2x2Mimo): enabling MCS0-MSC3 and MCS8-MCS11 on the accesspoint (select "advanced data rate", and "configured" to be able to do that) will limit TX/RX rate to 120MBps; do bandwidth tests and observe TX/RX CCQ in the wireless status or registration-tab until it reaches 90-100% and optimise the MCS settings. source is: http://en.wikipedia.org/wiki/IEEE_802.1 ... Data_rates Other settings that have influence to the jitter are: "adaptive noise immunity: none", "DFS mode:none", "HT guard interval: any"

One example for ping jitter of a 7km Link (sxt-sextant some trees in the line of sight)
with TDMA period size: 3ms and up to 240Mbps rates, , CCQ about 60%
tcp-download-rate (windows xp to a webserver): 1.2MByte/s:
UDP-Bandwidth test: TX/RX:80/55MBit/s
sent=100 received=100 packet-loss=0% min-rtt=2ms avg-rtt=15ms max-rtt=71ms

this is for TDMA period size: 1ms and 120Mbps rate, CCQ>95%
tcp-download-rate: 4MByte/s
UDP-Bandwidth test: TX/RX:60/50MBit/s
sent=100 received=100 packet-loss=0% min-rtt=2ms avg-rtt=5ms max-rtt=23ms
I use V5.22 on the Accesspoint an V5.26 on the client.

Quite interesting: the UDP-Rates are higher, but the TCP performance is much lower on a not optimised link. I should mention that I did not modify the tcp settings (rwin or similar) on the machines.

Leo.

See the smokeping below of the optimized link.
* at Thu 09:00 I switched from TDMA 3ms to 1ms
* at Thu 16:00 I changed the MCS settings

Leo.

The MCS Settings for 120MBit/s
Leo

This a very interest thread but I find it difficult to understand when LOS or freznel zone issues exist, that tweaking setting is attempted? Is it not best to have no LOS or Freznel zone issues first and then tweak setting.
Also having a good test set is very important and installed to ensure that variables which could effect results are minimized

leonix: I'm not using nv2, just straight 802.11n -- so I don't think any of the TDMA stuff applies to my situation.
n21roadie: my access-point and laptop are separated by maybe 5m with line of site -- so I don't think there are any fresnel zone issues (and if there were, they would apply equally to the UniFi setup).

Does anyone have any other ideas?

leonix: I'm not using nv2, just straight 802.11n -- so I don't think any of the TDMA stuff applies to my situation.
n21roadie: my access-point and laptop are separated by maybe 5m with line of site -- so I don't think there are any fresnel zone issues (and if there were, they would apply equally to the UniFi setup).

Does anyone have any other ideas?

The MCS settings are NOT NV2 related, this is also useful for pure 802.11n Links. And yes, sure first try to get your link clean (no obstacles, no interferences by local transmitters). However the solution above will also improve good link!
6.10 should also improve wireless timings, did anybody here already tested this?

M.

leonix,

Do you think that reducing MCS rates will increase throughput over my 5m link?
(Also: my CCQ is near 100%).

I'm running v6.10.

leonix,

Do you think that reducing MCS rates will increase throughput over my 5m link?
(Also: my CCQ is near 100%).

I'm running v6.10.

No. In case of CCQ~100 changing the MCS settings will not improve it. (maybe they changed the behaviour of automatic MCS selection in 6.10).

M.

Try reducing output power. For 5m away you don't need 27dB and even if it looks OK on your router it may cause reciever overload on your laptop.... also make sure you are using AES not TKIP for WPA2.

JF.

Thanks JanezFord.
My Tx/Rx signal strength for the laptop (as per the registration page) is -66dBm.
I tried turning down the tx power (by setting the antenna gain to 13dBi [when I actually have stubby omni antennas (5dBi?)]), no change.
Also I'm using the correct encryption.

I double-checked the CCQ: in fact it sits around 100% when not transmitting much, but when I do the iperf test:
- during the AP->STA part of the iperf test, it sits above 85%
- during the STA->AP part of the iperf test, it jumps around between 15% and 80%

I also looked at the "Rate/LastMeasured" part of the registration tab -- it only lists up to HT40-7. Is it always like this? Am I never getting MCS8+?

Borax: The CCQ during a full and massive Bandwidth test is important, the one in idle state not. More than HT40-7 is not possible, and because it is using a very sensitive modulation scheme in that case - it will do often several retransmits - which will improve the jitter of the packet latency. Just try it, set the MCS to a lower value.

Leo.

Hi Leo,

Why do you say that more than HT40-7 is not possible? My understanding is that MCS-8 uses the same modulation scheme as MCS-0, but with two spatial streams.

I tried playing with the MCS settings as you suggest. I'm not sure how to modify it as per your image:
- the web interface won't let me change anything
- (I don't have a windows machine for winbox)
- In the CLI:
it won't let me change ht-supported-mcs ("failure: supported MCS set does not include all basic")
I can change ht-basic-mcs (does not improve performance [and my client says it's still getting the MCS values I excluded]

The notation is quite confusing

HT40-7 is MCS15 (40MHz BW) -> 300Mbps (using guard interval "any" which is 400ns and 2 spatial streams)
..
HT40-0 is MCS08 (40MHz BW) -> 15Mbps

and for small bandwidth:
HT20-7 is MCS15 (20MHz BW) -> 144.4Mbps
..
HT20-0 is MCS08 (20MHz BW) -> 7.2Mbps

In case the system is using only one spatial stream (e.g. the other one is noisy) the system will use MCS00 - MCS07 settings. (in fact You have to allow MCS0-MCS07 also - otherwise the link is not working reliable, but
it will report HT40-7 in this case also!

Leo.

Thanks leonix, I hadn't understood the notation.

Bad throughput can also be a problem of wrong settings of the ethernet interfaces of the wireless routers!
http://forum.mikrotik.com/viewtopic.php ... 50#p462068

Leo.