summary | shortlog | log | commit | commitdiff | tree
raw | patch | inline | side by side (parent: 32755ef)
raw | patch | inline | side by side (parent: 32755ef)
author | Florian Forster <ff@octo.it> | |
Sun, 28 Sep 2014 09:17:26 +0000 (02:17 -0700) | ||
committer | Florian Forster <ff@octo.it> | |
Sun, 28 Sep 2014 09:20:56 +0000 (02:20 -0700) |
src/mans/oping.pod | patch | blob | history | |
src/oping.c | patch | blob | history |
diff --git a/src/mans/oping.pod b/src/mans/oping.pod
index b6ce0b2b7f35d12355cf4dbc1e222ab15ecbaec7..19d45042f6d237c451fb4649b6d75bce60bd6368 100644 (file)
--- a/src/mans/oping.pod
+++ b/src/mans/oping.pod
I<noping only> B<-u> forces UTF-8 output, B<-U> disables UTF-8 output. If
neither is given, the codeset is automatically determined from the locale.
+=item B<-P> I<percent>
+
+Configures the latency percentile to report. I<percent> must be a number
+between zero and 100, exclusively in both cases. In general, defaults to B<95>.
+If B<-c> is given and a number less than 20, this would be the same as the
+maximum. In this case the default is chosen so that it excludes the maximum,
+e.g. if B<-cE<nbsp>5> is given, the default is I<80>.
+
+The calculated percentile has roughly millisecond precision. If precision is of
+importance, read on for a more detailed explanation. In order to calculate the
+percentile without keeping all replies in memory, I<oping> divides the
+I<interval> (the B<-i> option) in 1000 "buckets". Each bucket counts the number
+of packets received in the represented time. That means that the precision
+decreases if the interval is increased, because each bucket represents a larger
+(fraction of the) response time. The code looks for the first bucket
+representing at least I<percent> responses and returns the upper-bound latency
+represented by that bucket. Since the represented percentage may be larger than
+the configured percentile, this algorithm I<overestimes> the actual percentile
+by at most 1000th of I<interval>.
+
=item B<-Z> I<percent>
If any hosts have a drop rate higher than I<percent>, where I<percent> is a
diff --git a/src/oping.c b/src/oping.c
index d53edd87ffc2285e34fc7567e83ff4889fcf3dcd..3a04fcd1ecfe7a1529e7cff44dca9b161a1083bf 100644 (file)
--- a/src/oping.c
+++ b/src/oping.c
double latency_total;
double latency_total_square;
+/* 1000 + one "infinity" bucket. */
+#define OPING_HISTOGRAM_BUCKETS 1001
+ uint32_t *latency_histogram;
+ size_t latency_histogram_size;
+
#if USE_NCURSES
WINDOW *window;
#endif
static int opt_count = -1;
static int opt_send_ttl = 64;
static uint8_t opt_send_qos = 0;
+#define OPING_DEFAULT_PERCENTILE 95.0
+static double opt_percentile = -1.0;
static double opt_exit_status_threshold = 1.0;
#if USE_NCURSES
static int opt_utf8 = 0;
ret->latency_total = 0.0;
ret->latency_total_square = 0.0;
+ ret->latency_histogram_size = (size_t) OPING_HISTOGRAM_BUCKETS;
+ ret->latency_histogram = calloc (ret->latency_histogram_size,
+ sizeof (*ret->latency_histogram));
+
#if USE_NCURSES
ret->window = NULL;
#endif
}
#endif
+ free (context->latency_histogram);
+ context->latency_histogram = NULL;
+
free (context);
} /* }}} void context_destroy */
return (ctx->latency_total / num_total);
} /* }}} double context_get_average */
+static double context_get_percentile (ping_context_t *ctx, /* {{{ */
+ double percentile)
+{
+ double threshold = percentile / 100.0;
+ uint32_t accumulated[ctx->latency_histogram_size];
+ double ratios[ctx->latency_histogram_size];
+ double index_to_ms_factor;
+ uint32_t num;
+ size_t i;
+
+ if (ctx->latency_histogram == NULL)
+ return (NAN);
+
+ accumulated[0] = ctx->latency_histogram[0];
+ for (i = 1; i < ctx->latency_histogram_size; i++)
+ accumulated[i] = accumulated[i - 1]
+ + ctx->latency_histogram[i];
+ num = accumulated[ctx->latency_histogram_size - 1];
+
+ for (i = 0; i < ctx->latency_histogram_size; i++)
+ {
+ ratios[i] = ((double) accumulated[i]) / ((double) num);
+ if (ratios[i] >= threshold)
+ break;
+ }
+
+ if (i >= ctx->latency_histogram_size)
+ return (NAN);
+ else if (i == (ctx->latency_histogram_size - 1))
+ return (INFINITY);
+
+ index_to_ms_factor = (1000.0 * opt_interval) / (ctx->latency_histogram_size - 1);
+
+ /* Multiply with i+1, because we're interested in the _upper_ bound of
+ * each bucket. */
+ return (index_to_ms_factor * ((double) (i + 1)));
+} /* }}} double context_get_percentile */
+
static double context_get_stddev (ping_context_t *ctx) /* {{{ */
{
double num_total;
#if USE_NCURSES
" -u / -U force / disable UTF-8 output\n"
#endif
+ " -P percent Report the n'th percentile of latency\n"
" -Z percent Exit with non-zero exit status if more than this percentage of\n"
" probes timed out. (default: never)\n"
while (1)
{
- optchar = getopt (argc, argv, "46c:hi:I:t:Q:f:D:Z:"
+ optchar = getopt (argc, argv, "46c:hi:I:t:Q:f:D:Z:P:"
#if USE_NCURSES
"uU"
#endif
int new_count;
new_count = atoi (optarg);
if (new_count > 0)
+ {
opt_count = new_count;
+
+ if ((opt_percentile < 0.0) && (opt_count < 20))
+ opt_percentile = 100.0 * (opt_count - 1) / opt_count;
+ }
else
fprintf(stderr, "Ignoring invalid count: %s\n",
optarg);
opt_interval = new_interval;
}
break;
+
case 'I':
{
if (opt_srcaddr != NULL)
set_opt_send_qos (optarg);
break;
+ case 'P':
+ {
+ double new_percentile;
+ new_percentile = atof (optarg);
+ if (isnan (new_percentile)
+ || (new_percentile < 0.1)
+ || (new_percentile > 100.0))
+ fprintf (stderr, "Ignoring invalid percentile: %s\n",
+ optarg);
+ else
+ opt_percentile = new_percentile;
+ }
+ break;
+
#if USE_NCURSES
case 'u':
opt_utf8 = 2;
}
}
+ if (opt_percentile <= 0.0)
+ opt_percentile = OPING_DEFAULT_PERCENTILE;
+
return (optind);
} /* }}} read_options */
@@ -811,14 +887,19 @@ static int update_stats_from_context (ping_context_t *ctx, pingobj_iter_t *iter)
{
double average;
double deviation;
+ double percentile;
average = context_get_average (ctx);
deviation = context_get_stddev (ctx);
+ percentile = context_get_percentile (ctx, opt_percentile);
mvwprintw (ctx->window, /* y = */ 2, /* x = */ 2,
- "rtt min/avg/max/sdev = %.3f/%.3f/%.3f/%.3f ms",
+ "rtt min/avg/%.0f%%/max/sdev = "
+ "%.3f/%.3f/%.0f/%.3f/%.3f ms\n",
+ opt_percentile,
ctx->latency_min,
average,
+ percentile,
ctx->latency_max,
deviation);
}
@@ -1015,6 +1096,30 @@ static int post_sleep_hook (__attribute__((unused)) pingobj_t *ping) /* {{{ */
} /* }}} int post_sleep_hook */
#endif
+static void update_context (ping_context_t *context, double latency) /* {{{ */
+{
+ size_t bucket;
+
+ context->req_rcvd++;
+ context->latency_total += latency;
+ context->latency_total_square += (latency * latency);
+
+ if ((context->latency_max < 0.0) || (context->latency_max < latency))
+ context->latency_max = latency;
+ if ((context->latency_min < 0.0) || (context->latency_min > latency))
+ context->latency_min = latency;
+
+ if (context->latency_histogram == NULL)
+ return;
+
+ /* latency is in ms, opt_interval is in s. */
+ bucket = (size_t) ((latency * (context->latency_histogram_size - 1))
+ / (1000.0 * opt_interval));
+ if (bucket >= context->latency_histogram_size)
+ bucket = context->latency_histogram_size - 1;
+ context->latency_histogram[bucket]++;
+} /* }}} void update_context */
+
static void update_host_hook (pingobj_iter_t *iter, /* {{{ */
__attribute__((unused)) int index)
{
context->req_sent++;
if (latency > 0.0)
{
- context->req_rcvd++;
- context->latency_total += latency;
- context->latency_total_square += (latency * latency);
-
- if ((context->latency_max < 0.0) || (context->latency_max < latency))
- context->latency_max = latency;
- if ((context->latency_min < 0.0) || (context->latency_min > latency))
- context->latency_min = latency;
+ update_context (context, latency);
#if USE_NCURSES
if (has_colors () == TRUE)
{
double average;
double deviation;
+ double percentile;
average = context_get_average (context);
deviation = context_get_stddev (context);
+ percentile = context_get_percentile (context, opt_percentile);
- printf ("rtt min/avg/max/sdev = %.3f/%.3f/%.3f/%.3f ms\n",
+ printf ("rtt min/avg/%.0f%%/max/sdev = "
+ "%.3f/%.3f/%.0f/%.3f/%.3f ms\n",
+ opt_percentile,
context->latency_min,
average,
+ percentile,
context->latency_max,
deviation);
}