Skip to contentSkip to conversation

ath9k: remove ath9k_rate_control

RFC In review Felix Fietkau 9 participants 37 messages Feb 8, 2013 – Mar 3, 2013 Qualcomm Atheros ATH9K wireless driver

lore
Share & tools
lore
#0 lore raw

RFC 1/1 ath9k: remove ath9k_rate_control

In many tests throughput with ath9k_rate_control has been shown to be worse than with minstrel_ht (the mac8021 default RC).

This module also has some other problems, like starting to use the highest possible rate early in the connection, causing problems with reliability of connection/authentication attempts.

It also has a much more limited search space, ignoring many potentially useful rates, caused by the design decision to operate on a sorted rate set with the assumption that higher rates are always more unreliable than lower rates. In some scenarios this assumption is not true, and this can cause it to fall back to a really bad rate.

minstrel_ht has been tested extensively in AP and client mode by lots of users (mostly in the OpenWrt project, where it has been the default for years).

The only advantage that ath9k_rate_control previously had over minstrel_ht was the support for using CCK rates as fall back in case MCS rates got too bad, but this has now also been taken care of.

---
 drivers/net/wireless/ath/ath9k/Kconfig  |    8 -
 drivers/net/wireless/ath/ath9k/Makefile |    1 -
 drivers/net/wireless/ath/ath9k/debug.h  |    1 -
 drivers/net/wireless/ath/ath9k/hw.c     |   15 +-
 drivers/net/wireless/ath/ath9k/hw.h     |    4 +-
 drivers/net/wireless/ath/ath9k/init.c   |   17 +-
 drivers/net/wireless/ath/ath9k/rc.c     | 1501 -------------------------------
 drivers/net/wireless/ath/ath9k/rc.h     |  248 -----
 drivers/net/wireless/ath/ath9k/xmit.c   |    9 +-
 9 files changed, 10 insertions(+), 1794 deletions(-)
 delete mode 100644 drivers/net/wireless/ath/ath9k/rc.c
 delete mode 100644 drivers/net/wireless/ath/ath9k/rc.h
drivers/net/wireless/ath/ath9k/Kconfig +0 -8
index 17507dc..037d919 100644
@@ -92,14 +92,6 @@ config ATH9K_MAC_DEBUG
 	  This option enables collection of statistics for Rx/Tx status
 	  data and some other MAC related statistics
 
-config ATH9K_RATE_CONTROL
-	bool "Atheros ath9k rate control"
-	depends on ATH9K
-	default y
-	---help---
-	  Say Y, if you want to use the ath9k specific rate control
-	  module instead of minstrel_ht.
-
 config ATH9K_HTC
        tristate "Atheros HTC based wireless cards support"
        depends on USB && MAC80211
drivers/net/wireless/ath/ath9k/Makefile +0 -1
index 2ad8f94..655e21d 100644
@@ -8,7 +8,6 @@ ath9k-y +=	beacon.o \
 		antenna.o
 
 ath9k-$(CONFIG_ATH9K_BTCOEX_SUPPORT) += mci.o
-ath9k-$(CONFIG_ATH9K_RATE_CONTROL) += rc.o
 ath9k-$(CONFIG_ATH9K_PCI) += pci.o
 ath9k-$(CONFIG_ATH9K_AHB) += ahb.o
 ath9k-$(CONFIG_ATH9K_DEBUGFS) += debug.o
drivers/net/wireless/ath/ath9k/debug.h +0 -1
index 410d6d8..06b5760 100644
@@ -18,7 +18,6 @@
 #define DEBUG_H
 
 #include "hw.h"
-#include "rc.h"
 #include "dfs_debug.h"
 
 struct ath_txq;
drivers/net/wireless/ath/ath9k/hw.c +3 -12
index 42cf3c7..6491973 100644
@@ -21,7 +21,6 @@
 
 #include "hw.h"
 #include "hw-ops.h"
-#include "rc.h"
 #include "ar9003_mac.h"
 #include "ar9003_mci.h"
 #include "ar9003_phy.h"
@@ -230,7 +229,7 @@ u32 ath9k_hw_reverse_bits(u32 val, u32 n)
 }
 
 u16 ath9k_hw_computetxtime(struct ath_hw *ah,
-			   u8 phy, int kbps,
+			   bool cck, int kbps,
 			   u32 frameLen, u16 rateix,
 			   bool shortPreamble)
 {
@@ -239,15 +238,13 @@ u16 ath9k_hw_computetxtime(struct ath_hw *ah,
 	if (kbps == 0)
 		return 0;
 
-	switch (phy) {
-	case WLAN_RC_PHY_CCK:
+	if (cck) {
 		phyTime = CCK_PREAMBLE_BITS + CCK_PLCP_BITS;
 		if (shortPreamble)
 			phyTime >>= 1;
 		numBits = frameLen << 3;
 		txTime = CCK_SIFS_TIME + phyTime + ((numBits * 1000) / kbps);
-		break;
-	case WLAN_RC_PHY_OFDM:
+	} else {
 		if (ah->curchan && IS_CHAN_QUARTER_RATE(ah->curchan)) {
 			bitsPerSymbol =	(kbps * OFDM_SYMBOL_TIME_QUARTER) / 1000;
 			numBits = OFDM_PLCP_BITS + (frameLen << 3);
@@ -270,12 +267,6 @@ u16 ath9k_hw_computetxtime(struct ath_hw *ah,
 			txTime = OFDM_SIFS_TIME + OFDM_PREAMBLE_TIME
 				+ (numSymbols * OFDM_SYMBOL_TIME);
 		}
-		break;
-	default:
-		ath_err(ath9k_hw_common(ah),
-			"Unknown phy %u (rate ix %u)\n", phy, rateix);
-		txTime = 0;
-		break;
 	}
 
 	return txTime;
drivers/net/wireless/ath/ath9k/hw.h +3 -1
index 784e81c..24f5b70 100644
@@ -22,6 +22,8 @@
 #include <linux/io.h>
 #include <linux/firmware.h>
 
+struct ath_softc;
+
 #include "mac.h"
 #include "ani.h"
 #include "eeprom.h"
@@ -1013,7 +1015,7 @@ void ath9k_hw_write_array(struct ath_hw *ah, const struct ar5416IniArray *array,
 			  int column, unsigned int *writecnt);
 u32 ath9k_hw_reverse_bits(u32 val, u32 n);
 u16 ath9k_hw_computetxtime(struct ath_hw *ah,
-			   u8 phy, int kbps,
+			   bool cck, int kbps,
 			   u32 frameLen, u16 rateix, bool shortPreamble);
 void ath9k_hw_get_channel_centers(struct ath_hw *ah,
 				  struct ath9k_channel *chan,
drivers/net/wireless/ath/ath9k/init.c +1 -16
index af932c9..c86cc70 100644
@@ -809,10 +809,6 @@ void ath9k_set_hw_capab(struct ath_softc *sc, struct ieee80211_hw *hw)
 	sc->ant_rx = hw->wiphy->available_antennas_rx;
 	sc->ant_tx = hw->wiphy->available_antennas_tx;
 
-#ifdef CONFIG_ATH9K_RATE_CONTROL
-	hw->rate_control_algorithm = "ath9k_rate_control";
-#endif
-
 	if (sc->sc_ah->caps.hw_caps & ATH9K_HW_CAP_2GHZ)
 		hw->wiphy->bands[IEEE80211_BAND_2GHZ] =
 			&sc->sbands[IEEE80211_BAND_2GHZ];
@@ -952,19 +948,11 @@ static int __init ath9k_init(void)
 {
 	int error;
 
-	/* Register rate control algorithm */
-	error = ath_rate_control_register();
-	if (error != 0) {
-		pr_err("Unable to register rate control algorithm: %d\n",
-		       error);
-		goto err_out;
-	}
-
 	error = ath_pci_init();
 	if (error < 0) {
 		pr_err("No PCI devices found, driver not installed\n");
 		error = -ENODEV;
-		goto err_rate_unregister;
+		goto err_out;
 	}
 
 	error = ath_ahb_init();
@@ -978,8 +966,6 @@ static int __init ath9k_init(void)
  err_pci_exit:
 	ath_pci_exit();
 
- err_rate_unregister:
-	ath_rate_control_unregister();
  err_out:
 	return error;
 }
@@ -990,7 +976,6 @@ static void __exit ath9k_exit(void)
 	is_ath9k_unloaded = true;
 	ath_ahb_exit();
 	ath_pci_exit();
-	ath_rate_control_unregister();
 	pr_info("%s: Driver unloaded\n", dev_info);
 }
 module_exit(ath9k_exit);
drivers/net/wireless/ath/ath9k/rc.c deleted +0 -1501

1,501 lines not on this page · raw message

drivers/net/wireless/ath/ath9k/rc.h deleted +0 -248

248 lines not on this page · raw message

drivers/net/wireless/ath/ath9k/xmit.c +3 -6
index feacaaf..b32fda8 100644
@@ -1008,8 +1008,7 @@ static void ath_buf_set_rate(struct ath_softc *sc, struct ath_buf *bf,
 	info->rtscts_rate = fi->rtscts_rate;
 
 	for (i = 0; i < 4; i++) {
-		bool is_40, is_sgi, is_sp;
-		int phy;
+		bool is_40, is_sgi, is_sp, cck = false;
 
 		if (!rates[i].count || (rates[i].idx < 0))
 			continue;
@@ -1050,9 +1049,7 @@ static void ath_buf_set_rate(struct ath_softc *sc, struct ath_buf *bf,
 		rate = &sc->sbands[tx_info->band].bitrates[rates[i].idx];
 		if ((tx_info->band == IEEE80211_BAND_2GHZ) &&
 		    !(rate->flags & IEEE80211_RATE_ERP_G))
-			phy = WLAN_RC_PHY_CCK;
-		else
-			phy = WLAN_RC_PHY_OFDM;
+			cck = true;
 
 		info->rates[i].Rate = rate->hw_value;
 		if (rate->hw_value_short) {
@@ -1069,7 +1066,7 @@ static void ath_buf_set_rate(struct ath_softc *sc, struct ath_buf *bf,
 					ah->txchainmask, info->rates[i].Rate);
 
 		info->rates[i].PktDuration = ath9k_hw_computetxtime(sc->sc_ah,
-			phy, rate->bitrate * 100, len, rix, is_sp);
+			cck, rate->bitrate * 100, len, rix, is_sp);
 	}
 
 	/* For AR5416 - RTS cannot be followed by a frame larger than 8K */
Signature · 1 line
1.8.0.2
#1 lore raw
Quoted 13 lines · Felix Fietkau

In many tests throughput with ath9k_rate_control has been shown to be worse than with minstrel_ht (the mac8021 default RC).

This module also has some other problems, like starting to use the highest possible rate early in the connection, causing problems with reliability of connection/authentication attempts.

It also has a much more limited search space, ignoring many potentially useful rates, caused by the design decision to operate on a sorted rate set with the assumption that higher rates are always more unreliable than lower rates. In some scenarios this assumption is not true, and this can cause it to fall back to a really bad rate.

Quoted 7 lines

minstrel_ht has been tested extensively in AP and client mode by lots of users (mostly in the OpenWrt project, where it has been the default for years).

The only advantage that ath9k_rate_control previously had over minstrel_ht was the support for using CCK rates as fall back in case MCS rates got too bad, but this has now also been taken care of.

Various rates are marked as invalid in the ath9k rate control module. I don't know all the reasons, but the recommendation to disable certain rates came from the internal algorithms team. How will this be handled ?

Sujith

#2 lore raw
Quoted 18 lines · Sujith Manoharan
Quoted 17 lines · Felix Fietkau

In many tests throughput with ath9k_rate_control has been shown to be worse than with minstrel_ht (the mac8021 default RC).

This module also has some other problems, like starting to use the highest possible rate early in the connection, causing problems with reliability of connection/authentication attempts.

It also has a much more limited search space, ignoring many potentially useful rates, caused by the design decision to operate on a sorted rate set with the assumption that higher rates are always more unreliable than lower rates. In some scenarios this assumption is not true, and this can cause it to fall back to a really bad rate.

minstrel_ht has been tested extensively in AP and client mode by lots of users (mostly in the OpenWrt project, where it has been the default for years).

Quoted 7 lines
Quoted 3 lines

The only advantage that ath9k_rate_control previously had over minstrel_ht was the support for using CCK rates as fall back in case MCS rates got too bad, but this has now also been taken care of.

Various rates are marked as invalid in the ath9k rate control module. I don't know all the reasons, but the recommendation to disable certain rates came from the internal algorithms team. How will this be handled ?

I'm pretty sure it's because the algorithm isn't designed to handle more rates. It's a step-up/step-down type algorithm, and those typically only work with a sorted rateset, preferably duplicate free.

Minstrel_ht selects rates in a different way that doesn't have the same limitations, so it can use the full rateset.

- Felix

#3 lore raw
Quoted 6 lines · Felix Fietkau

I'm pretty sure it's because the algorithm isn't designed to handle more rates. It's a step-up/step-down type algorithm, and those typically only work with a sorted rateset, preferably duplicate free.

Minstrel_ht selects rates in a different way that doesn't have the same limitations, so it can use the full rateset.

Makes sense.

Sujith

Acked-by on v1 1/1
#4 lore raw
In reply to Felix Fietkau
Quoted 13 lines · Felix Fietkau

In many tests throughput with ath9k_rate_control has been shown to be worse than with minstrel_ht (the mac8021 default RC).

This module also has some other problems, like starting to use the highest possible rate early in the connection, causing problems with reliability of connection/authentication attempts.

It also has a much more limited search space, ignoring many potentially useful rates, caused by the design decision to operate on a sorted rate set with the assumption that higher rates are always more unreliable than lower rates. In some scenarios this assumption is not true, and this can cause it to fall back to a really bad rate.

Quoted 3 lines

minstrel_ht has been tested extensively in AP and client mode by lots of users (mostly in the OpenWrt project, where it has been the default for years).

I don't think managed mode would have received much testing in OpenWRT.

Quoted 5 lines

The only advantage that ath9k_rate_control previously had over minstrel_ht was the support for using CCK rates as fall back in case MCS rates got too bad, but this has now also been taken care of.

Sujith

#5 lore raw
Quoted 14 lines · Sujith Manoharan
Quoted 13 lines · Felix Fietkau

In many tests throughput with ath9k_rate_control has been shown to be worse than with minstrel_ht (the mac8021 default RC).

This module also has some other problems, like starting to use the highest possible rate early in the connection, causing problems with reliability of connection/authentication attempts.

It also has a much more limited search space, ignoring many potentially useful rates, caused by the design decision to operate on a sorted rate set with the assumption that higher rates are always more unreliable than lower rates. In some scenarios this assumption is not true, and this can cause it to fall back to a really bad rate.

Quoted 5 lines
Quoted 3 lines

minstrel_ht has been tested extensively in AP and client mode by lots of users (mostly in the OpenWrt project, where it has been the default for years).

I don't think managed mode would have received much testing in OpenWRT.

I know quite a few people that use directional links with 4-addr AP/Station. I also know a few people that put up some devices that roam between multiple APs. Obviously it doesn't receive as much testing as AP mode, but there are some active users stress testing it.

- Felix

#6 lore raw

Could you share some of the test results? I know with ChromeOS when we switched to 3.4, the first thing that tipped us off to the fact that we had inadvertently switch to minstrel was that our RvR throughput values had dropped significantly. Has anyone done an RvR comparison between minstrel and the ath9k internal rate control lately?

--
Paul

Quoted 31 lines · Felix Fietkau
Quoted 14 lines · Sujith Manoharan
Quoted 13 lines · Felix Fietkau

In many tests throughput with ath9k_rate_control has been shown to be worse than with minstrel_ht (the mac8021 default RC).

This module also has some other problems, like starting to use the highest possible rate early in the connection, causing problems with reliability of connection/authentication attempts.

It also has a much more limited search space, ignoring many potentially useful rates, caused by the design decision to operate on a sorted rate set with the assumption that higher rates are always more unreliable than lower rates. In some scenarios this assumption is not true, and this can cause it to fall back to a really bad rate.

Quoted 5 lines
Quoted 3 lines

minstrel_ht has been tested extensively in AP and client mode by lots of users (mostly in the OpenWrt project, where it has been the default for years).

I don't think managed mode would have received much testing in OpenWRT.

I know quite a few people that use directional links with 4-addr AP/Station. I also know a few people that put up some devices that roam between multiple APs. Obviously it doesn't receive as much testing as AP mode, but there are some active users stress testing it.

- Felix

--
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

#7 lore raw
Quoted 6 lines · Paul Stewart

Could you share some of the test results? I know with ChromeOS when we switched to 3.4, the first thing that tipped us off to the fact that we had inadvertently switch to minstrel was that our RvR throughput values had dropped significantly. Has anyone done an RvR comparison between minstrel and the ath9k internal rate control lately?

Are you sure it was minstrel_ht, or could it also have been minstrel (the legacy version)? I don't have any recent RvR test results, however on every single link that I tested both rate control modules on (leaving everything else unchanged), minstrel_ht got slightly better results. Additionally, I occasionally get asked for help on debugging low throughput issues, and many of the people asking me have reported that simply switching to minstrel_ht fixed these issues in their tests.

I would also like to see some real RvR tests comparing both, preferably before this gets merged, which is why I sent it as RFC.

- Felix

#8 lore raw
Quoted 15 lines · Felix Fietkau
Quoted 6 lines · Paul Stewart

Could you share some of the test results? I know with ChromeOS when we switched to 3.4, the first thing that tipped us off to the fact that we had inadvertently switch to minstrel was that our RvR throughput values had dropped significantly. Has anyone done an RvR comparison between minstrel and the ath9k internal rate control lately?

Are you sure it was minstrel_ht, or could it also have been minstrel (the legacy version)? I don't have any recent RvR test results, however on every single link that I tested both rate control modules on (leaving everything else unchanged), minstrel_ht got slightly better results. Additionally, I occasionally get asked for help on debugging low throughput issues, and many of the people asking me have reported that simply switching to minstrel_ht fixed these issues in their tests.

Quoted 2 lines

I would also like to see some real RvR tests comparing both, preferably before this gets merged, which is why I sent it as RFC.

Paul, curious if you've run any tests?

Luis

#9 lore raw

Why don't we just move the ath9k rate control code to another mac80211 rate control module, so it's available to other devices?

It may not be a bad idea to keep it around as a reference and for people to do comparisons against. It just seems silly to have a rate control framework and then not use it..

Adrian

#10 lore raw
Quoted 6 lines · Adrian Chadd

Why don't we just move the ath9k rate control code to another mac80211 rate control module, so it's available to other devices?

It may not be a bad idea to keep it around as a reference and for people to do comparisons against. It just seems silly to have a rate control framework and then not use it..

I don't see the point in keeping that algorithm around. It has known design flaws that make it perform poorly in several scenarios, and fixing some of the flaws require basically rewriting it.

- Felix

#11 lore raw
Quoted 3 lines · Felix Fietkau

I don't see the point in keeping that algorithm around. It has known design flaws that make it perform poorly in several scenarios, and fixing some of the flaws require basically rewriting it.

I ran some comparison throughput tests between minstrel_ht and ath9k RC and minstrel showed lower numbers (~10Mbps difference). This is in a clean environment (OTA). IMO, we can switch the default to minstrel_ht, but keep the ath9k RC until the perf. gap is fixed. I think this was Google's concern too, a drop in the average throughput. This was in the STA->AP direction.

In the AP->STA direction (AP: DB120, STA: AR9280), the numbers were really low. The highest that I could see was ~165 Mbps, while with WNDR3700 (stock FW), the average throughput was ~185 Mbps and an occasional high of 190 and above.

Sujith

#12 lore raw
Quoted 10 lines · Sujith Manoharan
Quoted 3 lines · Felix Fietkau

I don't see the point in keeping that algorithm around. It has known design flaws that make it perform poorly in several scenarios, and fixing some of the flaws require basically rewriting it.

I ran some comparison throughput tests between minstrel_ht and ath9k RC and minstrel showed lower numbers (~10Mbps difference). This is in a clean environment (OTA). IMO, we can switch the default to minstrel_ht, but keep the ath9k RC until the perf. gap is fixed. I think this was Google's concern too, a drop in the average throughput. This was in the STA->AP direction.

What was the distance and what was the rate used?

Quoted 3 lines

In the AP->STA direction (AP: DB120, STA: AR9280), the numbers were really low. The highest that I could see was ~165 Mbps, while with WNDR3700 (stock FW), the average throughput was ~185 Mbps and an occasional high of 190 and above.

Comparing against stock FW (especially with different devices) is not useful, it could be anything.

- Felix

#13 lore raw
Quoted 1 line · Felix Fietkau

What was the distance and what was the rate used?

The station was about a meter away from the AP. I don't have the rcstats, debugfs was compiled out. I'll rerun the test and post the results.

Sujith

#14 lore raw
In reply to Felix Fietkau
Quoted 1 line · Felix Fietkau

What was the distance and what was the rate used?

A distance of about a meter between the AP and STA.

AP:  DB120 - running an internal build.
STA: WB222 - wireless-testing HEAD.

Signal is about -43 dBm.
Channel 40, HT40-

I've attached sample runs with minstrel_ht and ath9k-rc. With minstrel_ht, the numbers fluctuate between 180 and 190 Mbps and the average throughput varies each time (185, 186 etc.). With ath9k-rc, it's 195+ always.

Sujith

root@linux-test ~# iperf -i1 -c 172.16.1.222 -t 120 -P 4
------------------------------------------------------------
Client connecting to 172.16.1.222, TCP port 5001 TCP window size: 22.9 KByte (default)
------------------------------------------------------------

[  3] local 172.16.1.111 port 36815 connected with 172.16.1.222 port 5001
[  4] local 172.16.1.111 port 36816 connected with 172.16.1.222 port 5001
[  5] local 172.16.1.111 port 36817 connected with 172.16.1.222 port 5001
[  6] local 172.16.1.111 port 36818 connected with 172.16.1.222 port 5001
     19:[SUM]    0.0- 1.0 sec  22.8 MBytes   191 Mbits/sec
     24:[SUM]    1.0- 2.0 sec  23.0 MBytes   193 Mbits/sec
     29:[SUM]    2.0- 3.0 sec  23.6 MBytes   198 Mbits/sec
     34:[SUM]    3.0- 4.0 sec  22.9 MBytes   192 Mbits/sec
     39:[SUM]    4.0- 5.0 sec  23.2 MBytes   195 Mbits/sec
     44:[SUM]    5.0- 6.0 sec  22.6 MBytes   190 Mbits/sec
     49:[SUM]    6.0- 7.0 sec  23.2 MBytes   195 Mbits/sec
     54:[SUM]    7.0- 8.0 sec  23.8 MBytes   199 Mbits/sec
     59:[SUM]    8.0- 9.0 sec  23.1 MBytes   194 Mbits/sec
     64:[SUM]    9.0-10.0 sec  23.4 MBytes   196 Mbits/sec
     69:[SUM]   10.0-11.0 sec  23.0 MBytes   193 Mbits/sec
     74:[SUM]   11.0-12.0 sec  23.2 MBytes   195 Mbits/sec
     79:[SUM]   12.0-13.0 sec  22.8 MBytes   191 Mbits/sec
     84:[SUM]   13.0-14.0 sec  23.0 MBytes   193 Mbits/sec
     89:[SUM]   14.0-15.0 sec  23.1 MBytes   194 Mbits/sec
     94:[SUM]   15.0-16.0 sec  23.0 MBytes   193 Mbits/sec
     99:[SUM]   16.0-17.0 sec  22.5 MBytes   189 Mbits/sec
    104:[SUM]   17.0-18.0 sec  23.2 MBytes   195 Mbits/sec
    109:[SUM]   18.0-19.0 sec  23.4 MBytes   196 Mbits/sec
    114:[SUM]   19.0-20.0 sec  23.4 MBytes   196 Mbits/sec
    119:[SUM]   20.0-21.0 sec  22.9 MBytes   192 Mbits/sec
    124:[SUM]   21.0-22.0 sec  22.6 MBytes   190 Mbits/sec
    129:[SUM]   22.0-23.0 sec  23.4 MBytes   196 Mbits/sec
    134:[SUM]   23.0-24.0 sec  23.0 MBytes   193 Mbits/sec
    139:[SUM]   24.0-25.0 sec  22.6 MBytes   190 Mbits/sec
    144:[SUM]   25.0-26.0 sec  23.5 MBytes   197 Mbits/sec
    149:[SUM]   26.0-27.0 sec  23.1 MBytes   194 Mbits/sec
    154:[SUM]   27.0-28.0 sec  23.2 MBytes   195 Mbits/sec
    159:[SUM]   28.0-29.0 sec  23.8 MBytes   199 Mbits/sec
    164:[SUM]   29.0-30.0 sec  23.4 MBytes   196 Mbits/sec
    169:[SUM]   30.0-31.0 sec  22.8 MBytes   191 Mbits/sec
    174:[SUM]   31.0-32.0 sec  23.5 MBytes   197 Mbits/sec
    179:[SUM]   32.0-33.0 sec  22.8 MBytes   191 Mbits/sec
    184:[SUM]   33.0-34.0 sec  23.1 MBytes   194 Mbits/sec
    189:[SUM]   34.0-35.0 sec  23.2 MBytes   195 Mbits/sec
    194:[SUM]   35.0-36.0 sec  23.2 MBytes   195 Mbits/sec
    199:[SUM]   36.0-37.0 sec  23.1 MBytes   194 Mbits/sec
    204:[SUM]   37.0-38.0 sec  23.0 MBytes   193 Mbits/sec
    209:[SUM]   38.0-39.0 sec  23.2 MBytes   195 Mbits/sec
    214:[SUM]   39.0-40.0 sec  23.8 MBytes   199 Mbits/sec
    219:[SUM]   40.0-41.0 sec  23.0 MBytes   193 Mbits/sec
    224:[SUM]   41.0-42.0 sec  23.4 MBytes   196 Mbits/sec
    229:[SUM]   42.0-43.0 sec  23.5 MBytes   197 Mbits/sec
    234:[SUM]   43.0-44.0 sec  23.5 MBytes   197 Mbits/sec
    239:[SUM]   44.0-45.0 sec  22.6 MBytes   190 Mbits/sec
    244:[SUM]   45.0-46.0 sec  23.2 MBytes   195 Mbits/sec
    249:[SUM]   46.0-47.0 sec  23.2 MBytes   195 Mbits/sec
    254:[SUM]   47.0-48.0 sec  23.1 MBytes   194 Mbits/sec
    259:[SUM]   48.0-49.0 sec  24.1 MBytes   202 Mbits/sec
    264:[SUM]   49.0-50.0 sec  23.0 MBytes   193 Mbits/sec
    269:[SUM]   50.0-51.0 sec  23.4 MBytes   196 Mbits/sec
    274:[SUM]   51.0-52.0 sec  23.4 MBytes   196 Mbits/sec
    279:[SUM]   52.0-53.0 sec  23.4 MBytes   196 Mbits/sec
    284:[SUM]   53.0-54.0 sec  23.5 MBytes   197 Mbits/sec
    289:[SUM]   54.0-55.0 sec  23.2 MBytes   195 Mbits/sec
    294:[SUM]   55.0-56.0 sec  23.2 MBytes   195 Mbits/sec
    299:[SUM]   56.0-57.0 sec  23.2 MBytes   195 Mbits/sec
    304:[SUM]   57.0-58.0 sec  22.8 MBytes   191 Mbits/sec
    309:[SUM]   58.0-59.0 sec  23.6 MBytes   198 Mbits/sec
    314:[SUM]   59.0-60.0 sec  23.1 MBytes   194 Mbits/sec
    319:[SUM]   60.0-61.0 sec  22.9 MBytes   192 Mbits/sec
    324:[SUM]   61.0-62.0 sec  23.4 MBytes   196 Mbits/sec
    329:[SUM]   62.0-63.0 sec  23.4 MBytes   196 Mbits/sec
    334:[SUM]   63.0-64.0 sec  22.5 MBytes   189 Mbits/sec
    339:[SUM]   64.0-65.0 sec  23.8 MBytes   199 Mbits/sec
    344:[SUM]   65.0-66.0 sec  23.0 MBytes   193 Mbits/sec
    349:[SUM]   66.0-67.0 sec  24.1 MBytes   202 Mbits/sec
    354:[SUM]   67.0-68.0 sec  22.6 MBytes   190 Mbits/sec
    359:[SUM]   68.0-69.0 sec  23.0 MBytes   193 Mbits/sec
    364:[SUM]   69.0-70.0 sec  23.2 MBytes   195 Mbits/sec
    369:[SUM]   70.0-71.0 sec  23.2 MBytes   195 Mbits/sec
    374:[SUM]   71.0-72.0 sec  22.2 MBytes   187 Mbits/sec
    379:[SUM]   72.0-73.0 sec  23.1 MBytes   194 Mbits/sec
    384:[SUM]   73.0-74.0 sec  23.9 MBytes   200 Mbits/sec
    389:[SUM]   74.0-75.0 sec  22.9 MBytes   192 Mbits/sec
    394:[SUM]   75.0-76.0 sec  23.6 MBytes   198 Mbits/sec
    399:[SUM]   76.0-77.0 sec  23.1 MBytes   194 Mbits/sec
    404:[SUM]   77.0-78.0 sec  23.1 MBytes   194 Mbits/sec
    409:[SUM]   78.0-79.0 sec  23.8 MBytes   199 Mbits/sec
    414:[SUM]   79.0-80.0 sec  23.8 MBytes   199 Mbits/sec
    419:[SUM]   80.0-81.0 sec  23.1 MBytes   194 Mbits/sec
    424:[SUM]   81.0-82.0 sec  22.9 MBytes   192 Mbits/sec
    429:[SUM]   82.0-83.0 sec  23.4 MBytes   196 Mbits/sec
    434:[SUM]   83.0-84.0 sec  23.4 MBytes   196 Mbits/sec
    439:[SUM]   84.0-85.0 sec  23.5 MBytes   197 Mbits/sec
    444:[SUM]   85.0-86.0 sec  23.9 MBytes   200 Mbits/sec
    449:[SUM]   86.0-87.0 sec  22.9 MBytes   192 Mbits/sec
    454:[SUM]   87.0-88.0 sec  23.4 MBytes   196 Mbits/sec
    459:[SUM]   88.0-89.0 sec  23.4 MBytes   196 Mbits/sec
    464:[SUM]   89.0-90.0 sec  23.4 MBytes   196 Mbits/sec
    469:[SUM]   90.0-91.0 sec  22.9 MBytes   192 Mbits/sec
    474:[SUM]   91.0-92.0 sec  23.6 MBytes   198 Mbits/sec
    479:[SUM]   92.0-93.0 sec  23.1 MBytes   194 Mbits/sec
    484:[SUM]   93.0-94.0 sec  23.2 MBytes   195 Mbits/sec
    489:[SUM]   94.0-95.0 sec  23.2 MBytes   195 Mbits/sec
    494:[SUM]   95.0-96.0 sec  23.1 MBytes   194 Mbits/sec
    499:[SUM]   96.0-97.0 sec  22.9 MBytes   192 Mbits/sec
    504:[SUM]   97.0-98.0 sec  23.2 MBytes   195 Mbits/sec
    509:[SUM]   98.0-99.0 sec  23.4 MBytes   196 Mbits/sec
    514:[SUM]  99.0-100.0 sec  23.5 MBytes   197 Mbits/sec
    519:[SUM] 100.0-101.0 sec  24.0 MBytes   201 Mbits/sec
    524:[SUM] 101.0-102.0 sec  23.5 MBytes   197 Mbits/sec
    529:[SUM] 102.0-103.0 sec  23.0 MBytes   193 Mbits/sec
    534:[SUM] 103.0-104.0 sec  23.1 MBytes   194 Mbits/sec
    539:[SUM] 104.0-105.0 sec  23.4 MBytes   196 Mbits/sec
    544:[SUM] 105.0-106.0 sec  23.0 MBytes   193 Mbits/sec
    549:[SUM] 106.0-107.0 sec  23.5 MBytes   197 Mbits/sec
    554:[SUM] 107.0-108.0 sec  23.4 MBytes   196 Mbits/sec
    559:[SUM] 108.0-109.0 sec  23.0 MBytes   193 Mbits/sec
    564:[SUM] 109.0-110.0 sec  22.6 MBytes   190 Mbits/sec
    569:[SUM] 110.0-111.0 sec  23.8 MBytes   199 Mbits/sec
    574:[SUM] 111.0-112.0 sec  23.2 MBytes   195 Mbits/sec
    579:[SUM] 112.0-113.0 sec  23.6 MBytes   198 Mbits/sec
    584:[SUM] 113.0-114.0 sec  24.1 MBytes   202 Mbits/sec
    589:[SUM] 114.0-115.0 sec  22.4 MBytes   188 Mbits/sec
    594:[SUM] 115.0-116.0 sec  23.1 MBytes   194 Mbits/sec
    599:[SUM] 116.0-117.0 sec  23.4 MBytes   196 Mbits/sec
    604:[SUM] 117.0-118.0 sec  22.9 MBytes   192 Mbits/sec
    609:[SUM] 118.0-119.0 sec  22.9 MBytes   192 Mbits/sec
    617:[SUM] 119.0-120.0 sec  23.4 MBytes   196 Mbits/sec
    619:[SUM]   0.0-120.0 sec  2.72 GBytes   195 Mbits/sec

cat /sys/kernel/debug/ieee80211/phy2/netdev:wlan0/stations/00:03:7f:42:08:56/rc_stats

HT MCS Rate Success Retries XRetries PER

  HT20      0   6.5:          0          0          0          0
  HT40      0  13.5:          0          0          0          0
  HT40      1  27.5:          0          0          0          0
  HT40      2  40.5:          0          0          0          0
  HT40      3  54.0:          0          0          0          0
  HT40      4  81.5:          0          0          0          0
  HT40     11 108.0:          0          0          0          0
  HT40     12 162.0:          0          0          0          0
  HT40     13 216.0:          1          0          0          0
  HT40     14 243.0:          2          0          0          0
  HT40     15 270.0:        314         35          0          0
  HT40     15 300.0:      64236       6607          8          0

cat /sys/kernel/debug/ieee80211/phy2/ath9k/xmit

BE BK VI VO

MPDUs Queued:                2          0         0        53
MPDUs Completed:             2          0         0        53
MPDUs XRetried:              0          0         0         0
Aggregates:              64462          0         0         0
AMPDUs Queued HW:           46          0         0         0
AMPDUs Queued SW:      2018211          0         0         0
AMPDUs Completed:      2018257          0         0         0
AMPDUs Retried:          21058          0         0         0
AMPDUs XRetried:             0          0         0         0
FIFO Underrun:               0          0         0         0
TXOP Exceeded:               0          0         0         0
TXTIMER Expiry:              0          0         0         0
DESC CFG Error:              0          0         0         0
DATA Underrun:               0          0         0         0
DELIM Underrun:              0          0         0         0
TX-Pkts-All:           2018259          0         0        53
TX-Bytes-All:       3099857109          0         0      2440
HW-put-tx-buf:           64534          0         0        53
HW-tx-start:                 0          0         0         0
HW-tx-proc-desc:         64534          0         0        53
TX-Failed:                   0          0         0         0

root@linux-test ~# iperf -i1 -c 172.16.1.222 -t 120 -P 4
------------------------------------------------------------
Client connecting to 172.16.1.222, TCP port 5001 TCP window size: 22.9 KByte (default)
------------------------------------------------------------

[  5] local 172.16.1.111 port 36820 connected with 172.16.1.222 port 5001
[  4] local 172.16.1.111 port 36819 connected with 172.16.1.222 port 5001
[  3] local 172.16.1.111 port 36821 connected with 172.16.1.222 port 5001
[  6] local 172.16.1.111 port 36822 connected with 172.16.1.222 port 5001
    685:[SUM]    0.0- 1.0 sec  21.0 MBytes   176 Mbits/sec
    690:[SUM]    1.0- 2.0 sec  21.8 MBytes   182 Mbits/sec
    695:[SUM]    2.0- 3.0 sec  21.8 MBytes   182 Mbits/sec
    700:[SUM]    3.0- 4.0 sec  22.8 MBytes   191 Mbits/sec
    705:[SUM]    4.0- 5.0 sec  22.5 MBytes   189 Mbits/sec
    710:[SUM]    5.0- 6.0 sec  22.9 MBytes   192 Mbits/sec
    715:[SUM]    6.0- 7.0 sec  22.2 MBytes   187 Mbits/sec
    720:[SUM]    7.0- 8.0 sec  22.0 MBytes   185 Mbits/sec
    725:[SUM]    8.0- 9.0 sec  23.1 MBytes   194 Mbits/sec
    730:[SUM]    9.0-10.0 sec  21.9 MBytes   184 Mbits/sec
    735:[SUM]   10.0-11.0 sec  22.5 MBytes   189 Mbits/sec
    740:[SUM]   11.0-12.0 sec  22.1 MBytes   186 Mbits/sec
    745:[SUM]   12.0-13.0 sec  22.5 MBytes   189 Mbits/sec
    750:[SUM]   13.0-14.0 sec  23.0 MBytes   193 Mbits/sec
    755:[SUM]   14.0-15.0 sec  21.5 MBytes   180 Mbits/sec
    760:[SUM]   15.0-16.0 sec  22.5 MBytes   189 Mbits/sec
    765:[SUM]   16.0-17.0 sec  22.8 MBytes   191 Mbits/sec
    770:[SUM]   17.0-18.0 sec  22.8 MBytes   191 Mbits/sec
    775:[SUM]   18.0-19.0 sec  21.9 MBytes   184 Mbits/sec
    780:[SUM]   19.0-20.0 sec  21.1 MBytes   177 Mbits/sec
    785:[SUM]   20.0-21.0 sec  22.8 MBytes   191 Mbits/sec
    790:[SUM]   21.0-22.0 sec  22.5 MBytes   189 Mbits/sec
    795:[SUM]   22.0-23.0 sec  22.5 MBytes   189 Mbits/sec
    800:[SUM]   23.0-24.0 sec  22.6 MBytes   190 Mbits/sec
    805:[SUM]   24.0-25.0 sec  21.5 MBytes   180 Mbits/sec
    810:[SUM]   25.0-26.0 sec  22.1 MBytes   186 Mbits/sec
    815:[SUM]   26.0-27.0 sec  22.1 MBytes   186 Mbits/sec
    820:[SUM]   27.0-28.0 sec  21.6 MBytes   181 Mbits/sec
    825:[SUM]   28.0-29.0 sec  22.6 MBytes   190 Mbits/sec
    830:[SUM]   29.0-30.0 sec  23.0 MBytes   193 Mbits/sec
    835:[SUM]   30.0-31.0 sec  22.8 MBytes   191 Mbits/sec
    840:[SUM]   31.0-32.0 sec  22.6 MBytes   190 Mbits/sec
    845:[SUM]   32.0-33.0 sec  21.9 MBytes   184 Mbits/sec
    850:[SUM]   33.0-34.0 sec  21.6 MBytes   181 Mbits/sec
    855:[SUM]   34.0-35.0 sec  22.9 MBytes   192 Mbits/sec
    860:[SUM]   35.0-36.0 sec  22.0 MBytes   185 Mbits/sec
    865:[SUM]   36.0-37.0 sec  22.6 MBytes   190 Mbits/sec
    870:[SUM]   37.0-38.0 sec  22.2 MBytes   187 Mbits/sec
    875:[SUM]   38.0-39.0 sec  22.5 MBytes   189 Mbits/sec
    880:[SUM]   39.0-40.0 sec  21.9 MBytes   184 Mbits/sec
    885:[SUM]   40.0-41.0 sec  22.6 MBytes   190 Mbits/sec
    890:[SUM]   41.0-42.0 sec  22.2 MBytes   187 Mbits/sec
    895:[SUM]   42.0-43.0 sec  21.9 MBytes   184 Mbits/sec
    900:[SUM]   43.0-44.0 sec  22.4 MBytes   188 Mbits/sec
    905:[SUM]   44.0-45.0 sec  20.5 MBytes   172 Mbits/sec
    910:[SUM]   45.0-46.0 sec  22.4 MBytes   188 Mbits/sec
    915:[SUM]   46.0-47.0 sec  23.0 MBytes   193 Mbits/sec
    920:[SUM]   47.0-48.0 sec  21.8 MBytes   182 Mbits/sec
    925:[SUM]   48.0-49.0 sec  22.2 MBytes   187 Mbits/sec
    930:[SUM]   49.0-50.0 sec  22.4 MBytes   188 Mbits/sec
    935:[SUM]   50.0-51.0 sec  22.4 MBytes   188 Mbits/sec
    940:[SUM]   51.0-52.0 sec  23.1 MBytes   194 Mbits/sec
    945:[SUM]   52.0-53.0 sec  22.4 MBytes   188 Mbits/sec
    950:[SUM]   53.0-54.0 sec  22.6 MBytes   190 Mbits/sec
    955:[SUM]   54.0-55.0 sec  22.8 MBytes   191 Mbits/sec
    960:[SUM]   55.0-56.0 sec  22.4 MBytes   188 Mbits/sec
    965:[SUM]   56.0-57.0 sec  22.5 MBytes   189 Mbits/sec
    970:[SUM]   57.0-58.0 sec  23.0 MBytes   193 Mbits/sec
    975:[SUM]   58.0-59.0 sec  22.1 MBytes   186 Mbits/sec
    980:[SUM]   59.0-60.0 sec  21.8 MBytes   182 Mbits/sec
    985:[SUM]   60.0-61.0 sec  21.6 MBytes   181 Mbits/sec
    990:[SUM]   61.0-62.0 sec  22.6 MBytes   190 Mbits/sec
    995:[SUM]   62.0-63.0 sec  22.1 MBytes   186 Mbits/sec
   1000:[SUM]   63.0-64.0 sec  21.8 MBytes   182 Mbits/sec
   1005:[SUM]   64.0-65.0 sec  22.2 MBytes   187 Mbits/sec
   1010:[SUM]   65.0-66.0 sec  21.8 MBytes   182 Mbits/sec
   1015:[SUM]   66.0-67.0 sec  22.2 MBytes   187 Mbits/sec
   1020:[SUM]   67.0-68.0 sec  22.2 MBytes   187 Mbits/sec
   1025:[SUM]   68.0-69.0 sec  23.2 MBytes   195 Mbits/sec
   1030:[SUM]   69.0-70.0 sec  23.2 MBytes   195 Mbits/sec
   1035:[SUM]   70.0-71.0 sec  22.5 MBytes   189 Mbits/sec
   1040:[SUM]   71.0-72.0 sec  22.2 MBytes   187 Mbits/sec
   1045:[SUM]   72.0-73.0 sec  22.0 MBytes   185 Mbits/sec
   1050:[SUM]   73.0-74.0 sec  21.9 MBytes   184 Mbits/sec
   1055:[SUM]   74.0-75.0 sec  22.5 MBytes   189 Mbits/sec
   1060:[SUM]   75.0-76.0 sec  22.4 MBytes   188 Mbits/sec
   1065:[SUM]   76.0-77.0 sec  22.6 MBytes   190 Mbits/sec
   1070:[SUM]   77.0-78.0 sec  22.9 MBytes   192 Mbits/sec
   1075:[SUM]   78.0-79.0 sec  22.1 MBytes   186 Mbits/sec
   1080:[SUM]   79.0-80.0 sec  22.0 MBytes   185 Mbits/sec
   1085:[SUM]   80.0-81.0 sec  21.9 MBytes   184 Mbits/sec
   1090:[SUM]   81.0-82.0 sec  21.5 MBytes   180 Mbits/sec
   1095:[SUM]   82.0-83.0 sec  22.4 MBytes   188 Mbits/sec
   1100:[SUM]   83.0-84.0 sec  23.4 MBytes   196 Mbits/sec
   1105:[SUM]   84.0-85.0 sec  22.0 MBytes   185 Mbits/sec
   1110:[SUM]   85.0-86.0 sec  22.5 MBytes   189 Mbits/sec
   1115:[SUM]   86.0-87.0 sec  22.8 MBytes   191 Mbits/sec
   1120:[SUM]   87.0-88.0 sec  22.4 MBytes   188 Mbits/sec
   1125:[SUM]   88.0-89.0 sec  22.1 MBytes   186 Mbits/sec
   1130:[SUM]   89.0-90.0 sec  22.8 MBytes   191 Mbits/sec
   1135:[SUM]   90.0-91.0 sec  21.4 MBytes   179 Mbits/sec
   1140:[SUM]   91.0-92.0 sec  22.4 MBytes   188 Mbits/sec
   1145:[SUM]   92.0-93.0 sec  22.9 MBytes   192 Mbits/sec
   1150:[SUM]   93.0-94.0 sec  22.9 MBytes   192 Mbits/sec
   1155:[SUM]   94.0-95.0 sec  23.4 MBytes   196 Mbits/sec
   1160:[SUM]   95.0-96.0 sec  21.1 MBytes   177 Mbits/sec
   1165:[SUM]   96.0-97.0 sec  22.0 MBytes   185 Mbits/sec
   1170:[SUM]   97.0-98.0 sec  22.2 MBytes   187 Mbits/sec
   1175:[SUM]   98.0-99.0 sec  22.5 MBytes   189 Mbits/sec
   1180:[SUM]  99.0-100.0 sec  22.2 MBytes   187 Mbits/sec
   1185:[SUM] 100.0-101.0 sec  22.4 MBytes   188 Mbits/sec
   1190:[SUM] 101.0-102.0 sec  22.5 MBytes   189 Mbits/sec
   1195:[SUM] 102.0-103.0 sec  23.0 MBytes   193 Mbits/sec
   1200:[SUM] 103.0-104.0 sec  22.0 MBytes   185 Mbits/sec
   1205:[SUM] 104.0-105.0 sec  22.0 MBytes   185 Mbits/sec
   1210:[SUM] 105.0-106.0 sec  22.5 MBytes   189 Mbits/sec
   1215:[SUM] 106.0-107.0 sec  22.1 MBytes   186 Mbits/sec
   1220:[SUM] 107.0-108.0 sec  22.4 MBytes   188 Mbits/sec
   1225:[SUM] 108.0-109.0 sec  22.8 MBytes   191 Mbits/sec
   1230:[SUM] 109.0-110.0 sec  22.4 MBytes   188 Mbits/sec
   1235:[SUM] 110.0-111.0 sec  22.2 MBytes   187 Mbits/sec
   1240:[SUM] 111.0-112.0 sec  22.9 MBytes   192 Mbits/sec
   1245:[SUM] 112.0-113.0 sec  23.1 MBytes   194 Mbits/sec
   1250:[SUM] 113.0-114.0 sec  23.1 MBytes   194 Mbits/sec
   1255:[SUM] 114.0-115.0 sec  21.8 MBytes   182 Mbits/sec
   1260:[SUM] 115.0-116.0 sec  21.5 MBytes   180 Mbits/sec
   1265:[SUM] 116.0-117.0 sec  22.0 MBytes   185 Mbits/sec
   1270:[SUM] 117.0-118.0 sec  21.8 MBytes   182 Mbits/sec
   1275:[SUM] 118.0-119.0 sec  23.1 MBytes   194 Mbits/sec
   1283:[SUM] 119.0-120.0 sec  23.2 MBytes   195 Mbits/sec

1285:[SUM] 0.0-120.0 sec 2.62 GBytes 187 Mbits/sec

cat /sys/kernel/debug/ieee80211/phy3/netdev:wlan0/stations/00:03:7f:42:08:56/rc_stats

type         rate     throughput  ewma prob   this prob  retry   this succ/attempt   success    attempts
HT20/LGI     MCS0            6.6       99.9       100.0      2              0(  0)        86          91
HT20/LGI     MCS1           10.5       81.1       100.0      2              0(  0)        90          96
HT20/LGI     MCS2           17.5       91.8       100.0      2              0(  0)        88          91
HT20/LGI     MCS3           24.9       99.6       100.0      2              0(  0)        85          92
HT20/LGI     MCS4           36.0       99.2       100.0      2              0(  0)        88          92
HT20/LGI     MCS5           42.8       92.0       100.0      2              0(  0)        85          90
HT20/LGI     MCS6           48.2       94.0       100.0      2              0(  0)        85          95
HT20/LGI     MCS7           57.1       99.9       100.0      2              0(  0)        92          95
HT20/LGI     MCS8           12.9       99.5       100.0      2              0(  0)        89          93
HT20/LGI     MCS9           24.1       96.4       100.0      2              0(  0)        91          96
HT20/LGI     MCS10          36.3       99.9       100.0      2              0(  0)        86          91
HT20/LGI     MCS11          46.5       99.9       100.0      2              0(  0)        92          96
HT20/LGI     MCS12          60.3       91.1       100.0      2              0(  0)        89          93
HT20/LGI     MCS13          81.1       99.8       100.0      2              0(  0)        86          93
HT20/LGI     MCS14          90.0       99.9       100.0      2              0(  0)        91          96
HT20/LGI     MCS15          83.4       85.9       100.0      2              0(  0)        97          99
HT20/SGI     MCS0            7.3       99.6       100.0      2              0(  0)        87          93
HT20/SGI     MCS1           14.4       99.9       100.0      2              0(  0)        88          93
HT20/SGI     MCS2           20.9       99.3       100.0      2              0(  0)        88          93
HT20/SGI     MCS3           27.4       99.3       100.0      2              0(  0)        90          96
HT20/SGI     MCS4           37.9       95.2       100.0      2              0(  0)        83          92
HT20/SGI     MCS5           48.4       95.5       100.0      2              0(  0)        87          93
HT20/SGI     MCS6           55.6       99.6       100.0      2              0(  0)        93         100
HT20/SGI     MCS7           60.9       98.1       100.0      2              0(  0)        93          95
HT20/SGI     MCS8           13.7       95.3       100.0      2              0(  0)        82          90
HT20/SGI     MCS9           26.8       97.3       100.0      2              0(  0)        84          90
HT20/SGI     MCS10          39.8       99.9       100.0      2              0(  0)        90          91
HT20/SGI     MCS11          50.7       99.9       100.0      2              0(  0)        87          92
HT20/SGI     MCS12          71.1       98.9       100.0      2              0(  0)        85          89
HT20/SGI     MCS13          80.2       91.4       100.0      2              0(  0)        87          91
HT20/SGI     MCS14          97.0       99.9       100.0      2              0(  0)        92          97
HT20/SGI     MCS15          78.0       74.9         0.0      2              0(  0)       476         538
HT40/LGI     MCS0           13.5       99.8       100.0      2              0(  0)        90          95
HT40/LGI     MCS1           26.0       99.9       100.0      2              0(  0)        86          90
HT40/LGI     MCS2           36.1       96.6       100.0      2              0(  0)        89          92
HT40/LGI     MCS3           44.4       92.0       100.0      2              0(  0)        91          94
HT40/LGI     MCS4           61.7       90.7       100.0      2              0(  0)        86          93
HT40/LGI     MCS5           70.1       83.4       100.0      2              0(  0)        81          91
HT40/LGI     MCS6           85.9       92.0       100.0      2              0(  0)        91          96
HT40/LGI     MCS7          100.6       99.6       100.0      2              0(  0)        88          94
HT40/LGI     MCS8           25.5       97.9       100.0      2              0(  0)        83          92
HT40/LGI     MCS9           44.4       92.0       100.0      2              0(  0)        84          90
HT40/LGI     MCS10          67.7       99.5       100.0      2              0(  0)        81          91
HT40/LGI     MCS11          59.2       70.5         0.0      2              0(  0)        85          89
HT40/LGI     MCS12         105.8       92.0       100.0      3              0(  0)        90          93
HT40/LGI     MCS13         118.3       88.7       100.0      3              0(  0)        84          89
HT40/LGI     MCS14         145.4       97.4       100.0      4              0(  0)       768         899
HT40/LGI     MCS15         130.0       81.9       100.0      4              0(  0)     80856       88875
HT40/SGI     MCS0           14.9       99.9       100.0      2              0(  0)        86          91
HT40/SGI     MCS1           21.4       74.7         0.0      2              0(  0)        91          95
HT40/SGI     MCS2           41.1       99.9       100.0      2              0(  0)        85          91
HT40/SGI     MCS3           49.7       93.9       100.0      2              0(  0)        88          93
HT40/SGI     MCS4           55.4       74.8         0.0      2              0(  0)        88          92
HT40/SGI     MCS5           86.9       95.6       100.0      2              0(  0)        85          91
HT40/SGI     MCS6          100.9       99.9       100.0      2              0(  0)        86          94
HT40/SGI   P MCS7          106.9       98.4       100.0      5              0(  0)      4424        4980
HT40/SGI     MCS8           28.6       99.8       100.0      2              0(  0)        87          89
HT40/SGI     MCS9           45.1       85.3       100.0      2              0(  0)        85          91
HT40/SGI     MCS10          72.6       98.1       100.0      2              0(  0)        88          90
HT40/SGI     MCS11          89.9       98.9       100.0      2              0(  0)        80          88
HT40/SGI     MCS12         121.9       98.7       100.0      3              0(  0)        86          92
HT40/SGI     MCS13         132.4       94.0       100.0      3              0(  0)        85          89
HT40/SGI  t  MCS14         152.6       96.1       100.0      5              0(  0)     88243       96781
HT40/SGI T   MCS15         161.8       97.1       100.0      5              1(  1)   1760690     1945362

Total packet count:: ideal 1934621 lookaround 5901
Average A-MPDU length: 4.4

cat /sys/kernel/debug/ieee80211/phy3/ath9k/xmit

BE BK VI VO

MPDUs Queued:               10          0         0        47
MPDUs Completed:            10          0         0        47
MPDUs XRetried:              0          0         0         0
Aggregates:              63956          0         0         0
AMPDUs Queued HW:           99          0         0         0
AMPDUs Queued SW:      1940401          0         0         0
AMPDUs Completed:      1940500          0         0         0
AMPDUs Retried:          14536          0         0         0
AMPDUs XRetried:             0          0         0         0
FIFO Underrun:               0          0         0         0
TXOP Exceeded:               0          0         0         0
TXTIMER Expiry:              0          0         0         0
DESC CFG Error:              0          0         0         0
DATA Underrun:               0          0         0         0
DELIM Underrun:              0          0         0         0
TX-Pkts-All:           1940510          0         0        47
TX-Bytes-All:       2980293154          0         0      2296
HW-put-tx-buf:           70308          0         0        47
HW-tx-start:                 0          0         0         0
HW-tx-proc-desc:         70308          0         0        47
TX-Failed:                   0          0         0         0
#15 lore raw
Quoted 5 lines · Sujith Manoharan
Quoted 1 line · Felix Fietkau

What was the distance and what was the rate used?

A distance of about a meter between the AP and STA.

Quoted 8 lines
AP:  DB120 - running an internal build.
STA: WB222 - wireless-testing HEAD.

Signal is about -43 dBm.
Channel 40, HT40-

I've attached sample runs with minstrel_ht and ath9k-rc. With minstrel_ht, the numbers fluctuate between 180 and 190 Mbps and the average throughput varies each time (185, 186 etc.). With ath9k-rc, it's 195+ always.

We can run some tests through attenuators today.

I assume UDP would be good, or is TCP traffic preferred? (In our tests, we see significantly more throughput with UDP than TCP, even when we hacked on TCP ack delay).

Is wireless-testing the preferred tree to test against?

Thanks,
Ben

Quoted 3 lines

Sujith

Signature · 2 lines
Ben Greear <[email protected]>
Candela Technologies Inc  http://www.candelatech.com
#16 lore raw
Quoted 5 lines · Ben Greear

I assume UDP would be good, or is TCP traffic preferred? (In our tests, we see significantly more throughput with UDP than TCP, even when we hacked on TCP ack delay).

Is wireless-testing the preferred tree to test against?

Do both. Failures at higher MCS rates -> extra latency -> TCP backoff -> crappier throughput.

Adrian

#17 lore raw
In reply to Sujith Manoharan
Quoted 5 lines · Sujith Manoharan
Quoted 1 line · Felix Fietkau

What was the distance and what was the rate used?

A distance of about a meter between the AP and STA.

Quoted 8 lines
AP:  DB120 - running an internal build.
STA: WB222 - wireless-testing HEAD.

Signal is about -43 dBm.
Channel 40, HT40-

I've attached sample runs with minstrel_ht and ath9k-rc. With minstrel_ht, the numbers fluctuate between 180 and 190 Mbps and the average throughput varies each time (185, 186 etc.). With ath9k-rc, it's 195+ always.

Right. And the reason that it fluctuates is quite likely because at the higher MCS rates (6,7/14,15/22,23) the channel behaviour is not consistent and there's varying levels of packet loss, so things oscillate between a bunch of MCS rates there.

IIRC, the ath9k_rc algorithm is a lot more aggressive at sticking to a fixed, working MCS rate.

(190mbit TCP on 2-stream? Damn. Looks like I have a little more optimisation to do in FreeBSD. I'm only hitting 160/170mbit TCP on AR9280<->AR9280 and AR9280<->AR9287. But I hit 220-240mbit UDP.)

Thanks,

Adrian

#18 lore raw
Quoted 3 lines · Adrian Chadd

(190mbit TCP on 2-stream? Damn. Looks like I have a little more optimisation to do in FreeBSD. I'm only hitting 160/170mbit TCP on AR9280<->AR9280 and AR9280<->AR9287. But I hit 220-240mbit UDP.)

I used a AR9462 card for this test. With AR9280/AR9287 I get abysmal numbers with both rate control algorithms. Maybe something has regressed in ath9k for the AR9002 family...

Sujith

#19 lore raw
Quoted 8 lines · Sujith Manoharan
Quoted 3 lines · Adrian Chadd

(190mbit TCP on 2-stream? Damn. Looks like I have a little more optimisation to do in FreeBSD. I'm only hitting 160/170mbit TCP on AR9280<->AR9280 and AR9280<->AR9287. But I hit 220-240mbit UDP.)

I used a AR9462 card for this test. With AR9280/AR9287 I get abysmal numbers with both rate control algorithms. Maybe something has regressed in ath9k for the AR9002 family...

It's almost like we need some kind of lab space set aside for running regression tests with all of the ath9k hardware ,dating back to AR5416, with some automated testing rigs..

I have the hardware. Who wants to do it? :)

adrian

#20 lore raw

On 2013-03-01 21:40 -0800, Adrian Chadd wrote Sujith Manoharan:

AC>On 1 March 2013 18:19, Sujith Manoharan <[email protected]> wrote:
AC>> Adrian Chadd wrote:
AC>>> (190mbit TCP on 2-stream? Damn. Looks like I have a little more AC>>> optimisation to do in FreeBSD. I'm only hitting 160/170mbit TCP on AC>>> AR9280<->AR9280 and AR9280<->AR9287. But I hit 220-240mbit UDP.) AC>>
AC>> I used a AR9462 card for this test. With AR9280/AR9287 I get abysmal AC>> numbers with both rate control algorithms. Maybe something has regressed AC>> in ath9k for the AR9002 family...
AC>
AC>It's almost like we need some kind of lab space set aside for running AC>regression tests with all of the ath9k hardware ,dating back to AC>AR5416, with some automated testing rigs..

Think this is good idea, have same problem - no mo ~50 megabit on ar9002.

AC>I have the hardware. Who wants to do it? :)

C уважением With Best Regards
Георгиевский Юрий. Georgiewskiy Yuriy
+7 4872 711666 +7 4872 711666
факс +7 4872 711143 fax +7 4872 711143
Компания ООО "Ай Ти Сервис" IT Service Ltd

http://nkoort.ru                  http://nkoort.ru
JID: [email protected]            JID: [email protected]
YG129-RIPE                        YG129-RIPE
#21 lore raw
In reply to Sujith Manoharan
Quoted 5 lines · Sujith Manoharan
Quoted 1 line · Felix Fietkau

What was the distance and what was the rate used?

A distance of about a meter between the AP and STA.

Quoted 8 lines
AP:  DB120 - running an internal build.
STA: WB222 - wireless-testing HEAD.

Signal is about -43 dBm.
Channel 40, HT40-

I've attached sample runs with minstrel_ht and ath9k-rc. With minstrel_ht, the numbers fluctuate between 180 and 190 Mbps and the average throughput varies each time (185, 186 etc.). With ath9k-rc, it's 195+ always.

I just sent some minstrel_ht improvements to linux-wireless@ Please test if they fix the throughput delta in your setup. In my tests they visibly reduce the number of unnecessary fallbacks to lower rates.

Thanks,

- Felix

#22 lore raw
Quoted 3 lines · Felix Fietkau

I just sent some minstrel_ht improvements to linux-wireless@ Please test if they fix the throughput delta in your setup. In my tests they visibly reduce the number of unnecessary fallbacks to lower rates.

With the fixes, the avg. TX throughput crept up to 188 Mbps, but I ran ath9k-rc once again and obtained 197 Mbps. There were lots of 200+ Mbps max. throughput too.

This is not a synthetic benchmark with shield boxes, attenuators etc., just a simple over-the-air test with not much interference, so a gap of ~10 Mbps is somewhat surprising.

Sujith

#23 lore raw
In reply to Felix Fietkau
Quoted 7 lines · Felix Fietkau
Quoted 3 lines · Adrian Chadd

It may not be a bad idea to keep it around as a reference and for people to do comparisons against. It just seems silly to have a rate control framework and then not use it..

I don't see the point in keeping that algorithm around. It has known design flaws that make it perform poorly in several scenarios, and fixing some of the flaws require basically rewriting it.

The same could be said of PID...

Signature · 1 line
Bob Copeland %% www.bobcopeland.com
#24 lore raw
Quoted 10 lines · Bob Copeland
Quoted 7 lines · Felix Fietkau
Quoted 3 lines · Adrian Chadd

It may not be a bad idea to keep it around as a reference and for people to do comparisons against. It just seems silly to have a rate control framework and then not use it..

I don't see the point in keeping that algorithm around. It has known design flaws that make it perform poorly in several scenarios, and fixing some of the flaws require basically rewriting it.

The same could be said of PID...

I agree, we should remove that one as well.

- Felix

#25 lore raw
Quoted 2 lines · Felix Fietkau
Quoted 1 line

The same could be said of PID...

I agree, we should remove that one as well.

One of the advantages of having multiple RC modules - even if they're not longer optimal - is to keep the API honest. :-)

I still keep onoe/amrr working in FreeBSD's ath(4) driver, primarily to make sure that I don't change the API without thinking too much about what other rate control modules do, but also to provide a simpler example of how the API works.

Personally, I'd like to see more examples of rate control modules in LInux/FreeBSD, especially ones that start demanding more 802.11 state (ie, air-time QoS.)

Adrian

#26 lore raw
Quoted 5 lines · Adrian Chadd
Quoted 2 lines · Felix Fietkau
Quoted 1 line

The same could be said of PID...

I agree, we should remove that one as well.

Quoted 7 lines

One of the advantages of having multiple RC modules - even if they're not longer optimal - is to keep the API honest. :-)

I still keep onoe/amrr working in FreeBSD's ath(4) driver, primarily to make sure that I don't change the API without thinking too much about what other rate control modules do, but also to provide a simpler example of how the API works.

In that case I'd rather keep PID than the ath9k rate control. The ath9k rate control is a horrible example of how to use the rate control API, and fixing that is a waste of time in my opinion.
By the way, minstrel and minstrel_ht are two mostly separate implementations using the same API, except for the fact that minstrel_ht falls back to minstrel for legacy clients. So we already do have multiple examples here :)

Quoted 3 lines

Personally, I'd like to see more examples of rate control modules in LInux/FreeBSD, especially ones that start demanding more 802.11 state (ie, air-time QoS.)

Do you have any good ideas on what state information would be useful for a rate control to demand?

- Felix

#27 lore raw
Quoted 3 lines · Felix Fietkau

In that case I'd rather keep PID than the ath9k rate control. The ath9k rate control is a horrible example of how to use the rate control API, and fixing that is a waste of time in my opinion.

I acked the earlier RFC patch to remove the ath9k RC as I assumed that minstrel_ht would perform adequately, if not better. But, there appear to be areas in which the numbers given by minstrel_ht are sub-optimal and which can be improved if we have something to compare it with. The ath9k RC code is crap, no doubt, but I'd prefer to have it for some time in the tree - we can switch the default RC to minstrel_ht by default.

Can you also explain how the ath9k RC uses the mac80211 API horribly ? The algorithm might be terribly implemented and the internal code nasty, but how does it violate the API ?

Sujith

#28 lore raw
Quoted 11 lines · Sujith Manoharan
Quoted 3 lines · Felix Fietkau

In that case I'd rather keep PID than the ath9k rate control. The ath9k rate control is a horrible example of how to use the rate control API, and fixing that is a waste of time in my opinion.

I acked the earlier RFC patch to remove the ath9k RC as I assumed that minstrel_ht would perform adequately, if not better. But, there appear to be areas in which the numbers given by minstrel_ht are sub-optimal and which can be improved if we have something to compare it with. The ath9k RC code is crap, no doubt, but I'd prefer to have it for some time in the tree - we can switch the default RC to minstrel_ht by default.

Right, I will only submit this patch again once more performance tests have been run. My own tests and tests of other people that I know of so far have shown minstrel_ht to perform better than ath9k_rate_control, whereas in your tests ath9k_rate_control seems to perform better. I'd like to get some more details on your tests, e.g. raw numbers, rate control stats, info on how much the throughput fluctuates, etc.

Quoted 2 lines

Can you also explain how the ath9k RC uses the mac80211 API horribly ? The algorithm might be terribly implemented and the internal code nasty, but how does it violate the API ?

When I called it a horrible example, I was referring to the bad code quality, not API violations. The API violations are only minor, there are a few places left where the rate control code accesses struct ath_softc.

- Felix

#29 lore raw
In reply to Felix Fietkau
Quoted 6 lines · Felix Fietkau
Quoted 3 lines

Personally, I'd like to see more examples of rate control modules in LInux/FreeBSD, especially ones that start demanding more 802.11 state (ie, air-time QoS.)

Do you have any good ideas on what state information would be useful for a rate control to demand?

Well, there are already hooks to fake the TX time spent, right? Having both TX and RX busy times available to rate control so it can better schedule TX loads based on current RX workloads would be nice.

Having access to schedule which peer and how much to send to each peer would be nice. Stuff like "peer X only can have up to x ms in this WME class this round", so you don't have a busy, close peer monopolising the air. It also means you can start doing smart things with far away peers who retransmit a lot - they're likely tying up a lot of airtime.

None of this is new. It's just, you know, new to open source. :-)

Adrian

#30 lore raw
Quoted 13 lines · Adrian Chadd
Quoted 6 lines · Felix Fietkau
Quoted 3 lines

Personally, I'd like to see more examples of rate control modules in LInux/FreeBSD, especially ones that start demanding more 802.11 state (ie, air-time QoS.)

Do you have any good ideas on what state information would be useful for a rate control to demand?

Well, there are already hooks to fake the TX time spent, right? Having both TX and RX busy times available to rate control so it can better schedule TX loads based on current RX workloads would be nice.

Quoted 7 lines

Having access to schedule which peer and how much to send to each peer would be nice. Stuff like "peer X only can have up to x ms in this WME class this round", so you don't have a busy, close peer monopolising the air. It also means you can start doing smart things with far away peers who retransmit a lot - they're likely tying up a lot of airtime.

None of this is new. It's just, you know, new to open source. :-)

In my opinion this doesn't really belong into a rate control module. There should be a tx scheduling API to take care of this. Before I implement something like this, I plan on exposing all per-station driver queues to mac80211. This is necessary for a few other things anyway, e.g. unifying software aggregation logic and fixing its buffer management.

- Felix

#31 lore raw
Quoted 6 lines · Felix Fietkau
Quoted 6 lines

Having access to schedule which peer and how much to send to each peer would be nice. Stuff like "peer X only can have up to x ms in this WME class this round", so you don't have a busy, close peer monopolising the air. It also means you can start doing smart things with far away peers who retransmit a lot - they're likely tying up a lot of airtime.

Quoted 7 lines
Quoted 1 line

None of this is new. It's just, you know, new to open source. :-)

In my opinion this doesn't really belong into a rate control module. There should be a tx scheduling API to take care of this. Before I implement something like this, I plan on exposing all per-station driver queues to mac80211. This is necessary for a few other things anyway, e.g. unifying software aggregation logic and fixing its buffer management.

Sure, but then some more clever tricks end up being difficult to implement. For example, knowing if a client is tying up too much airtime at the given rate and whether to back them off for a bit. Or to use smaller aggregation limits for certain clients because your'e trying to be "fairer" when trying to keep latency low. That kind of thing.

I think "rate control" should likely be expanded to "tx scheduling" as a whole, rather than sitting as a separate thing that just selects the rate for a node who has already been chosen to transmit.

Adrian

#32 lore raw
Quoted 23 lines · Adrian Chadd
Quoted 6 lines · Felix Fietkau
Quoted 6 lines

Having access to schedule which peer and how much to send to each peer would be nice. Stuff like "peer X only can have up to x ms in this WME class this round", so you don't have a busy, close peer monopolising the air. It also means you can start doing smart things with far away peers who retransmit a lot - they're likely tying up a lot of airtime.

Quoted 7 lines
Quoted 1 line

None of this is new. It's just, you know, new to open source. :-)

In my opinion this doesn't really belong into a rate control module. There should be a tx scheduling API to take care of this. Before I implement something like this, I plan on exposing all per-station driver queues to mac80211. This is necessary for a few other things anyway, e.g. unifying software aggregation logic and fixing its buffer management.

Sure, but then some more clever tricks end up being difficult to implement. For example, knowing if a client is tying up too much airtime at the given rate and whether to back them off for a bit. Or to use smaller aggregation limits for certain clients because your'e trying to be "fairer" when trying to keep latency low. That kind of thing.

Quoted 3 lines

I think "rate control" should likely be expanded to "tx scheduling" as a whole, rather than sitting as a separate thing that just selects the rate for a node who has already been chosen to transmit.

Even with client airtime use, I still don't see how tx scheduling and rate control belong together. In my opinion, the rate selection should not be based on client airtime usage or the current load, as it can optimize for throughput/airtime ratio without it.

- Felix

#33 lore raw

Hi Felix,

Quoted 24 lines · Felix Fietkau
Quoted 23 lines · Adrian Chadd
Quoted 6 lines · Felix Fietkau
Quoted 6 lines

Having access to schedule which peer and how much to send to each peer would be nice. Stuff like "peer X only can have up to x ms in this WME class this round", so you don't have a busy, close peer monopolising the air. It also means you can start doing smart things with far away peers who retransmit a lot - they're likely tying up a lot of airtime.

Quoted 7 lines
Quoted 1 line

None of this is new. It's just, you know, new to open source. :-)

In my opinion this doesn't really belong into a rate control module. There should be a tx scheduling API to take care of this. Before I implement something like this, I plan on exposing all per-station driver queues to mac80211. This is necessary for a few other things anyway, e.g. unifying software aggregation logic and fixing its buffer management.

Sure, but then some more clever tricks end up being difficult to implement. For example, knowing if a client is tying up too much airtime at the given rate and whether to back them off for a bit. Or to use smaller aggregation limits for certain clients because your'e trying to be "fairer" when trying to keep latency low. That kind of thing.

Quoted 8 lines
Quoted 3 lines

I think "rate control" should likely be expanded to "tx scheduling" as a whole, rather than sitting as a separate thing that just selects the rate for a node who has already been chosen to transmit.

Even with client airtime use, I still don't see how tx scheduling and rate control belong together. In my opinion, the rate selection should not be based on client airtime usage or the current load, as it can optimize for throughput/airtime ratio without it.

Algorithm folks and Engineers had spent considerable time on ath9k rate control. Wouldn't be a great idea to remove it completely, We can have it optional. With lot of throughput tests ran over internally and with the test team verification,
it wouldn't be fair to throw it away.

Signature · 2 lines
thanks,
shafi
#34 lore raw
Quoted 36 lines · Mohammed Shafi

Hi Felix,

Quoted 24 lines · Felix Fietkau
Quoted 23 lines · Adrian Chadd
Quoted 6 lines · Felix Fietkau
Quoted 6 lines

Having access to schedule which peer and how much to send to each peer would be nice. Stuff like "peer X only can have up to x ms in this WME class this round", so you don't have a busy, close peer monopolising the air. It also means you can start doing smart things with far away peers who retransmit a lot - they're likely tying up a lot of airtime.

Quoted 7 lines
Quoted 1 line

None of this is new. It's just, you know, new to open source. :-)

In my opinion this doesn't really belong into a rate control module. There should be a tx scheduling API to take care of this. Before I implement something like this, I plan on exposing all per-station driver queues to mac80211. This is necessary for a few other things anyway, e.g. unifying software aggregation logic and fixing its buffer management.

Sure, but then some more clever tricks end up being difficult to implement. For example, knowing if a client is tying up too much airtime at the given rate and whether to back them off for a bit. Or to use smaller aggregation limits for certain clients because your'e trying to be "fairer" when trying to keep latency low. That kind of thing.

Quoted 8 lines
Quoted 3 lines

I think "rate control" should likely be expanded to "tx scheduling" as a whole, rather than sitting as a separate thing that just selects the rate for a node who has already been chosen to transmit.

Even with client airtime use, I still don't see how tx scheduling and rate control belong together. In my opinion, the rate selection should not be based on client airtime usage or the current load, as it can optimize for throughput/airtime ratio without it.

Quoted 5 lines

Algorithm folks and Engineers had spent considerable time on ath9k rate control. Wouldn't be a great idea to remove it completely, We can have it optional. With lot of throughput tests ran over internally and with the test team verification,
it wouldn't be fair to throw it away.

Regardless of how much time was spent tuning it, it still has a really bad design, bad implementation and a number of practical issues. It seems to be tuned entirely for artificial benchmarks in clean air. It also starts with a very high rate without having proven that it works. I don't think anybody is going to fix all of these issues, and even if somebody does, it would invalidate pretty much all of the tuning/testing that went into this code.

- Felix

#35 lore raw
Quoted 42 lines · Felix Fietkau
Quoted 36 lines · Mohammed Shafi

Hi Felix,

Quoted 24 lines · Felix Fietkau
Quoted 23 lines · Adrian Chadd
Quoted 6 lines · Felix Fietkau
Quoted 6 lines

Having access to schedule which peer and how much to send to each peer would be nice. Stuff like "peer X only can have up to x ms in this WME class this round", so you don't have a busy, close peer monopolising the air. It also means you can start doing smart things with far away peers who retransmit a lot - they're likely tying up a lot of airtime.

Quoted 7 lines
Quoted 1 line

None of this is new. It's just, you know, new to open source. :-)

In my opinion this doesn't really belong into a rate control module. There should be a tx scheduling API to take care of this. Before I implement something like this, I plan on exposing all per-station driver queues to mac80211. This is necessary for a few other things anyway, e.g. unifying software aggregation logic and fixing its buffer management.

Sure, but then some more clever tricks end up being difficult to implement. For example, knowing if a client is tying up too much airtime at the given rate and whether to back them off for a bit. Or to use smaller aggregation limits for certain clients because your'e trying to be "fairer" when trying to keep latency low. That kind of thing.

Quoted 8 lines
Quoted 3 lines

I think "rate control" should likely be expanded to "tx scheduling" as a whole, rather than sitting as a separate thing that just selects the rate for a node who has already been chosen to transmit.

Even with client airtime use, I still don't see how tx scheduling and rate control belong together. In my opinion, the rate selection should not be based on client airtime usage or the current load, as it can optimize for throughput/airtime ratio without it.

Quoted 5 lines

Algorithm folks and Engineers had spent considerable time on ath9k rate control. Wouldn't be a great idea to remove it completely, We can have it optional. With lot of throughput tests ran over internally and with the test team verification,
it wouldn't be fair to throw it away.

Quoted 8 lines

Regardless of how much time was spent tuning it, it still has a really bad design, bad implementation and a number of practical issues. It seems to be tuned entirely for artificial benchmarks in clean air. It also starts with a very high rate without having proven that it works. I don't think anybody is going to fix all of these issues, and even if somebody does, it would invalidate pretty much all of the tuning/testing that went into this code.

We can certainly question the design and implementation, but it was proven to work well and had been tested by more hands(even with propitiatory stuff). Even if some one thinks it as bad, we should still allow it as an option. For instance if the environment is pretty good/pretty bad, we should give the users an option to choose between the two, that's why we should retain it. Further its been tested internally, by customers and more folks. Its been also worked by some serious developers and Engineers.

Signature · 2 lines
thanks,
shafi
#36 lore raw
Quoted 52 lines · Mohammed Shafi
Quoted 42 lines · Felix Fietkau
Quoted 36 lines · Mohammed Shafi

Hi Felix,

Quoted 24 lines · Felix Fietkau
Quoted 23 lines · Adrian Chadd
Quoted 6 lines · Felix Fietkau
Quoted 6 lines

Having access to schedule which peer and how much to send to each peer would be nice. Stuff like "peer X only can have up to x ms in this WME class this round", so you don't have a busy, close peer monopolising the air. It also means you can start doing smart things with far away peers who retransmit a lot - they're likely tying up a lot of airtime.

Quoted 7 lines
Quoted 1 line

None of this is new. It's just, you know, new to open source. :-)

In my opinion this doesn't really belong into a rate control module. There should be a tx scheduling API to take care of this. Before I implement something like this, I plan on exposing all per-station driver queues to mac80211. This is necessary for a few other things anyway, e.g. unifying software aggregation logic and fixing its buffer management.

Sure, but then some more clever tricks end up being difficult to implement. For example, knowing if a client is tying up too much airtime at the given rate and whether to back them off for a bit. Or to use smaller aggregation limits for certain clients because your'e trying to be "fairer" when trying to keep latency low. That kind of thing.

Quoted 8 lines
Quoted 3 lines

I think "rate control" should likely be expanded to "tx scheduling" as a whole, rather than sitting as a separate thing that just selects the rate for a node who has already been chosen to transmit.

Even with client airtime use, I still don't see how tx scheduling and rate control belong together. In my opinion, the rate selection should not be based on client airtime usage or the current load, as it can optimize for throughput/airtime ratio without it.

Quoted 5 lines

Algorithm folks and Engineers had spent considerable time on ath9k rate control. Wouldn't be a great idea to remove it completely, We can have it optional. With lot of throughput tests ran over internally and with the test team verification,
it wouldn't be fair to throw it away.

Quoted 8 lines

Regardless of how much time was spent tuning it, it still has a really bad design, bad implementation and a number of practical issues. It seems to be tuned entirely for artificial benchmarks in clean air. It also starts with a very high rate without having proven that it works. I don't think anybody is going to fix all of these issues, and even if somebody does, it would invalidate pretty much all of the tuning/testing that went into this code.

Quoted 6 lines

We can certainly question the design and implementation, but it was proven to work well and had been tested by more hands(even with propitiatory stuff). Even if some one thinks it as bad, we should still allow it as an option. For instance if the environment is pretty good/pretty bad, we should give the users an option to choose between the two, that's why we should retain it.

I agree that we should keep it for a while, but we should probably change the default soon. As for giving *users* the option of choosing rate control based on the environment, I think it's a bad idea! Regular users can't really be expected to know enough about the details of rate control to make a meaningful decision, nor should they be. Simply being lazy and telling every user to just test which one works better for them is also a bad idea.

The default implementation needs to be good for all kinds of environments. minstrel_ht is already tuned to work quite well in tough environments with heavy interference, so closing the performance gap in clean environments should be quite easy. The main thing I need to get that done is good quality test feedback - something that I didn't get a lot of, neither from users, nor from QCA.

Quoted 1 line

Further its been tested internally, by customers and more folks.

Right, and a few of those customers complained about bad performance with ath9k, and also reported that switching to minstrel_ht fixed their issues :)
Good to know that it's been tested, but where are the results? Who is going to take care of the long standing *known* issues?
One of these issues (too high bitrate before rate feedback is available) led to somebody from Google hacking up a crappy workaround in wpa_supplicant that disables high bitrates during connect.

Saying that it has been tested does not necessarily imply that it's any good or that it works properly in all normal scenarios ;)

Quoted 1 line

Its been also worked by some serious developers and Engineers.

Well, I do consider myself a serious developer and engineer as well :)

- Felix