svn commit: r225402 - user/adrian/if_ath_tx/sys/dev/ath/ath_rate/sample

[Adrian Chadd]

Author: adrian
Date: Mon Sep  5 15:21:18 2011
New Revision: 225402
URL: http://svn.freebsd.org/changeset/base/225402

Log:
  As a temporary workaround, treat MCS error rates 5% higher than
  the best rate as "ok" when picking a best rate.
  
  This is all a hack to avoid having to implement the "correct"
  solution, whatever that is. For example, minstrel calculates
  a predicted A-MPDU throughput value based on the
  average A-MPDU sub-frame count and per-rate TX success probability.
  
  This, along with lowering sample_stats to 2 from 10 (ie, spend 2%
  of TX time trying to sample alternate rates) seems to keep my TCP
  TX mostly stable at around 120-130mbit for my testing.
  
  TODO:
  
  * The EWMA doesn't raise quickly enough and seems a bit wrong, eg:
  
  [ 0 MCS: 250]      329:329      (100%) (EWMA  63.6%) T       22 F    0 avg   407 last 19144
  
  .. how'd the EWMA come out to be 63% given a 100% success rate overall?
  It made an incorrect judgement early on and hasn't yet had enough time
  to correct it.
  
  The sampling of other MCS rates is also not optimal - it shouldn't try
  sampling all of them during the sample interval. Instead, come up with
  a more intelligent way of picking the MCS rates to sample.
  
  Finally, comparing MCS rates by number is incorrect. This really should
  compare the per-rate throughput value, as higher MCS rate != higher
  throughput.
  
  But, as said above, this is "good enough" for local testing.
  It shouldn't be merged into -HEAD like this.

Modified:
  user/adrian/if_ath_tx/sys/dev/ath/ath_rate/sample/sample.c

Modified: user/adrian/if_ath_tx/sys/dev/ath/ath_rate/sample/sample.c
==============================================================================
--- user/adrian/if_ath_tx/sys/dev/ath/ath_rate/sample/sample.c	Mon Sep  5 15:11:01 2011	(r225401)
+++ user/adrian/if_ath_tx/sys/dev/ath/ath_rate/sample/sample.c	Mon Sep  5 15:21:18 2011	(r225402)
@@ -206,13 +206,15 @@ pick_best_rate(struct ath_node *an, cons
 			continue;
 
 		/*
-		 * For HT, Don't use a bit rate that has a higher failure
-		 * rate than the current.
+		 * For HT, Don't use a bit rate that is much more
+		 * lossy than the best.
 		 *
-		 * XXX This isn't optimal!
+		 * XXX this isn't optimal; it's just designed to
+		 * eliminate rates that are going to be obviously
+		 * worse.
 		 */
 		if (an->an_node.ni_flags & IEEE80211_NODE_HT) {
-			if (best_rate_pct > pct)
+			if (best_rate_pct > (pct + 50))
 				continue;
 		}