Implement timeout- and load detection-logic

Also, elect to not do any SQM changes when there is no clear trend to be
determined.

diff --git a/__lagdetect.awk b/__lagdetect.awk
index 9c0f1b23484183595233c943be529289595a86cb..42778e5417beba5326aea996816982ea8f009ab6 100644 (file)
--- a/__lagdetect.awk
+++ b/__lagdetect.awk
@@ -23,7 +23,10 @@ BEGIN {
     ping_pprev[pn] = -1 # latency recorded in the previous cycle
     not_increasing_count[pn] = 0 # number of consecutive cycles with no (fudged) latency increase
   }
-  adjust_delta = 0
+  consec_timeouts = 0
+  adjust_ts_delta = 0
+  rx_rate_max = 100 # make something up
+  tx_rate_max = 100 # make something up
   update_bw()
   update_bw_min_delta = (0.02 * peer_count)
 }
@@ -37,6 +40,7 @@ BEGIN {
   lat=9999
   update_bw()
   # XXX TODO what now?
+  consec_timeouts++
   process_record(pn, bytes, lat)
   #print "TIMEOUT for " pn
 }
@@ -49,6 +53,7 @@ BEGIN {
   bytes=$4
   lat=$6
   update_bw()
+  consec_timeouts = 0
   process_record(pn, bytes, lat)
 }
 
@@ -76,6 +81,7 @@ function get_time() {
 
 
 function update_bw() {
+  # If the last update happened rather recently, do not compute new bw stats.
   if (! (update_ts())) {
     #print "too fast, skiping bw updates " ts - ts_old
     return
@@ -94,6 +100,14 @@ function update_bw() {
   rx_rate=(rx_delta / (ts_delta))
   tx_delta=(tx - tx_old)
   tx_rate=(tx_delta / (ts_delta))
+  if (tx_rate_max < tx_rate) {
+    tx_rate_max = tx_rate
+    print "# " ts " new peak tx_rate=" tx_rate_max
+  }
+  if (rx_rate_max < rx_rate) {
+    rx_rate_max = rx_rate
+    print "# " ts " new peak rx_rate=" rx_rate_max
+  }
   # printf("# " ts " (Kbyte/s) rx_rate=%-12.2f tx_rate=%-12.2f ts_delta=%-1.2f\n", (rx_rate / 1024), (tx_rate / 1024), ts_delta)
 }
 
@@ -106,14 +120,14 @@ function process_record(pn, bytes, lat) {
 
   if (bytes == 0 && lat == 9999) {
     print "# WARN: TIMEOUT: " pn
-    return 1
+    # XXX TODO: what to do now?
   }
 
   #printf("%s %db in %.2fms\n", pn, bytes, lat)
   update_pingstats(pn, lat)
 
   if (have_baseline[pn]) {
-    adjust_bw(pn, lat)
+    adjust_sqm(pn, lat)
   }
 
   ping_pprev[pn]=ping_prev[pn]
@@ -121,36 +135,55 @@ function process_record(pn, bytes, lat) {
 }
 
 
-function adjust_bw(peername, latency) {
-  adjust_delta = (ts - adjust_old)
+function adjust_sqm(peername, latency) {
+  adjust_ts_delta = (ts - adjust_old)
+
+  # Assume links are clogged, so we short-circuit a bandwidth decrease
+  if (consec_timeouts > ((1.5 * peer_count) + 2)) {
+    # print "# " ts " too many TIMEOUTS in a row, decreasing SQM bw"
+    set_bw(IFACE_RX, 5000)
+    set_bw(IFACE_TX, 5000)
+    return
+  }
 
   # do not try to set bw more than once a second
-  if (adjust_delta < 1.0) {
-    # print "looping too fast, skipping BW update adjust_delta=" adjust_delta
+  if (adjust_ts_delta < 1.0) {
+    # print "looping too fast, skipping SQM update adjust_ts_delta=" adjust_ts_delta
     return
   }
 
+  # XXX this logic needs fixing
+  if ( rx_rate < (rx_rate_max * 0.1) || tx_rate < (tx_rate_max * 0.1) ) {
+    # print "line does not appear to be loaded, skipping SQM update"
+    return
+  }
+
+  # Try to determine the latency trend over the last few samples
   last2avg = (((ping_prev[peername] + latency) / 2.0))
   thresh_fudged = (FUDGE + (ping_avgs[peername] * THRESHOLD))
   if (last2avg > thresh_fudged ) {
     if (latency > ping_prev[peername] && latency > ping_pprev[peername]) {
-      print ts " " adjust_delta " CONSIDER BW DECREASE # " peername " lat=" latency " prev=" ping_prev[peername] " pprev=" ping_pprev[peername] " avg=" ping_avgs[peername] " min=" ping_min[peername]
+      # print ts " " adjust_ts_delta " CONSIDER BW DECREASE # " peername " lat=" latency " prev=" ping_prev[peername] " pprev=" ping_pprev[peername] " avg=" ping_avgs[peername] " min=" ping_min[peername]
       set_bw(IFACE_TX, 5000)
       set_bw(IFACE_RX, 5000)
       # print "set adjust_old=" ts
       adjust_old = ts
     }
     not_increasing_count[peername]=0
-  } else {
+  } else if ((latency < FUDGE + ping_avgs[peername]) &&
+             latency < ping_prev[peername] &&
+            latency < ping_pprev[peername] &&
+            ping_prev[peername] < ping_pprev[peername]) {
     if (not_increasing_count[peername] > (5 * PINGSLOTS)) {
-      print ts " " adjust_delta " CONSIDER BW INCREASE # " peername " lat=" latency " prev=" ping_prev[peername] " pprev=" ping_pprev[peername] " avg=" ping_avgs[peername] " min=" ping_min[peername]
+      # print ts " " adjust_ts_delta " CONSIDER BW INCREASE # " peername " lat=" latency " prev=" ping_prev[peername] " pprev=" ping_pprev[peername] " avg=" ping_avgs[peername] " min=" ping_min[peername]
       not_increasing_count[peername]=0
       set_bw(IFACE_TX, 10000)
       set_bw(IFACE_RX, 10000)
-      # print "set adjust_old=" ts
       adjust_old = ts
     }
     not_increasing_count[peername]++
+  } else {
+    # print ts " NOOP - no latency trend determined"
   }
 }