From: Martin Sperl <kernel@martin.sperl.org>
Date: Tue, 6 May 2014 21:37:47 +0000 (+0000)
Subject: added interpolation for percentile
X-Git-Tag: v1.5.0-rc1~106^2
X-Git-Url: http://git.ipfire.org/?a=commitdiff_plain;h=refs%2Fpull%2F465%2Fhead;p=thirdparty%2Frrdtool-1.x.git

added interpolation for percentile
---

diff --git a/doc/rrdgraph_rpn.pod b/doc/rrdgraph_rpn.pod
index 80b427b9..fd1f3099 100644
--- a/doc/rrdgraph_rpn.pod
+++ b/doc/rrdgraph_rpn.pod
@@ -253,6 +253,22 @@ This also allows for extrapolation into the future (say a few days)
 - you may need to define the data series whit the optional start= parameter, so that 
 the source data series has enough data to provide prediction also at the beginning of a graph...
 
+The percentile can be between [-100:+100]. 
+The positive percentiles interpolates between values while the negative will take the closest.
+
+Example: you run 7 shifts with a window of 1800seconds. Assuming that the rrd-file
+has a step size of 300 seconds this means we have to do the percentile calculation
+based on a max of 42 distinct values (less if you got NAN). that means that in the 
+best case you get a step rate between values of 2.4 percent.
+so if you ask for the 99th percentile, then you would need to look at the 41.59th
+value. As we only have integers, either the 41st or the 42nd value.
+
+With the positive percentile a linear interpolation between the 2 values is done
+to get the effective value.
+
+The negative returns the closest value distance wise - so in the above case 42nd value,
+which is effectively returning the Percentile100 or the max of the previous 7 days in the window.
+
 Here an example, that will create a 10 day graph that also shows the 
 prediction 3 days into the future with its uncertainty value (as defined by avg+-4*sigma)
 This also shows if the prediction is exceeded at a certain point.
diff --git a/src/rrd_rpncalc.c b/src/rrd_rpncalc.c
index 80522717..1796e47b 100644
--- a/src/rrd_rpncalc.c
+++ b/src/rrd_rpncalc.c
@@ -855,7 +855,7 @@ short rpn_calc(
 		if (rpnp[rpi].op == OP_PREDICTPERC) {
 		    stackunderflow(1);
 		    percentile = rpnstack->s[--stptr];
-		    if ((percentile<0) || (percentile > 100)) {
+		    if (abs(percentile) > 100) {
 		        rrd_set_error("unsupported percentile: %f",percentile);
 			return -1;
 		    }
@@ -957,11 +957,19 @@ short rpn_calc(
 		        /* sort the numbers */
 		        qsort(extra,count,sizeof(double),rpn_compare_double);
 			/* get the percentile selected */
-			int idx=(int)round(percentile * ((float)count-1.0));
-			/* maybe we should also do an interpolation between the 2
-			 * neighboring fields, similar to what we do with MEDIAN 
-			 */
-			val = extra[idx];
+			double idxf = percentile * ((float)count-1.0);
+			if (percentile < 0) { /* take the next best */
+			    int idx = round(abs(idxf));
+			    val = extra[idx];
+			} else { /* interpolate */
+			    int idx = floor(idxf);
+			    double deltax = idxf - idx;
+			    val = extra[idx];
+			    if (deltax) { /* this check also handles the percentile=100 case */
+			      double deltay = extra[idx + 1] - extra[idx];
+			      val += deltay * deltax;
+			    }
+			}
 		    }
 		    break;
 		default: /* should not get here ... */