From: Stephen Hemminger <stephen@networkplumber.org>
To: dev@dpdk.org
Cc: "Stephen Hemminger" <stephen@networkplumber.org>,
"Tyler Retzlaff" <roretzla@linux.microsoft.com>,
"Morten Brørup" <mb@smartsharesystems.com>,
"Reshma Pattan" <reshma.pattan@intel.com>
Subject: [PATCH v5 3/9] latencystats: do not use floating point
Date: Mon, 22 Apr 2024 08:21:33 -0700 [thread overview]
Message-ID: <20240422152336.147553-4-stephen@networkplumber.org> (raw)
In-Reply-To: <20240422152336.147553-1-stephen@networkplumber.org>
The cycle counts do not need to be stored as floating point.
Instead keep track of latency in cycles, and convert to
nanoseconds when read.
Use scaled math with exponential Weighted Moving Average weight
of .25 to avoid use of floating point for that.
The average latency took too long to "warm up".
Do what RFC 6298 suggests and initialize on first sample.
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Tyler Retzlaff <roretzla@linux.microsoft.com>
Acked-by: Morten Brørup <mb@smartsharesystems.com>
---
lib/latencystats/rte_latencystats.c | 128 ++++++++++++++++------------
1 file changed, 73 insertions(+), 55 deletions(-)
diff --git a/lib/latencystats/rte_latencystats.c b/lib/latencystats/rte_latencystats.c
index 3152256066..2140ab87c4 100644
--- a/lib/latencystats/rte_latencystats.c
+++ b/lib/latencystats/rte_latencystats.c
@@ -4,6 +4,7 @@
#include <math.h>
+#include <rte_common.h>
#include <rte_string_fns.h>
#include <rte_mbuf_dyn.h>
#include <rte_log.h>
@@ -41,11 +42,14 @@ static uint64_t samp_intvl;
static uint64_t timer_tsc;
static uint64_t prev_tsc;
+#define LATENCY_AVG_SCALE 4
+#define LATENCY_JITTER_SCALE 16
+
struct rte_latency_stats {
- float min_latency; /**< Minimum latency in nano seconds */
- float avg_latency; /**< Average latency in nano seconds */
- float max_latency; /**< Maximum latency in nano seconds */
- float jitter; /** Latency variation */
+ uint64_t min_latency; /**< Minimum latency */
+ uint64_t avg_latency; /**< Average latency */
+ uint64_t max_latency; /**< Maximum latency */
+ uint64_t jitter; /** Latency variation */
rte_spinlock_t lock; /** Latency calculation lock */
};
@@ -61,31 +65,38 @@ static struct rxtx_cbs tx_cbs[RTE_MAX_ETHPORTS][RTE_MAX_QUEUES_PER_PORT];
struct latency_stats_nameoff {
char name[RTE_ETH_XSTATS_NAME_SIZE];
unsigned int offset;
+ unsigned int scale;
};
static const struct latency_stats_nameoff lat_stats_strings[] = {
- {"min_latency_ns", offsetof(struct rte_latency_stats, min_latency)},
- {"avg_latency_ns", offsetof(struct rte_latency_stats, avg_latency)},
- {"max_latency_ns", offsetof(struct rte_latency_stats, max_latency)},
- {"jitter_ns", offsetof(struct rte_latency_stats, jitter)},
+ {"min_latency_ns", offsetof(struct rte_latency_stats, min_latency), 1},
+ {"avg_latency_ns", offsetof(struct rte_latency_stats, avg_latency), LATENCY_AVG_SCALE},
+ {"max_latency_ns", offsetof(struct rte_latency_stats, max_latency), 1},
+ {"jitter_ns", offsetof(struct rte_latency_stats, jitter), LATENCY_JITTER_SCALE},
};
#define NUM_LATENCY_STATS (sizeof(lat_stats_strings) / \
sizeof(lat_stats_strings[0]))
-int32_t
-rte_latencystats_update(void)
+static void
+latencystats_collect(uint64_t values[])
{
unsigned int i;
- float *stats_ptr = NULL;
- uint64_t values[NUM_LATENCY_STATS] = {0};
- int ret;
+ const uint64_t *stats;
for (i = 0; i < NUM_LATENCY_STATS; i++) {
- stats_ptr = RTE_PTR_ADD(glob_stats,
- lat_stats_strings[i].offset);
- values[i] = floor(*stats_ptr / cycles_per_ns);
+ stats = RTE_PTR_ADD(glob_stats, lat_stats_strings[i].offset);
+ values[i] = floor(*stats / (cycles_per_ns * lat_stats_strings[i].scale));
}
+}
+
+int32_t
+rte_latencystats_update(void)
+{
+ uint64_t values[NUM_LATENCY_STATS];
+ int ret;
+
+ latencystats_collect(values);
ret = rte_metrics_update_values(RTE_METRICS_GLOBAL,
latency_stats_index,
@@ -97,16 +108,16 @@ rte_latencystats_update(void)
}
static void
-rte_latencystats_fill_values(struct rte_metric_value *values)
+rte_latencystats_fill_values(struct rte_metric_value *metrics)
{
+ uint64_t values[NUM_LATENCY_STATS];
unsigned int i;
- float *stats_ptr = NULL;
+
+ latencystats_collect(values);
for (i = 0; i < NUM_LATENCY_STATS; i++) {
- stats_ptr = RTE_PTR_ADD(glob_stats,
- lat_stats_strings[i].offset);
- values[i].key = i;
- values[i].value = floor(*stats_ptr / cycles_per_ns);
+ metrics[i].key = i;
+ metrics[i].value = values[i];
}
}
@@ -151,15 +162,9 @@ calc_latency(uint16_t pid __rte_unused,
void *_ __rte_unused)
{
unsigned int i;
- uint64_t now;
- float latency;
- static float prev_latency;
- /*
- * Alpha represents degree of weighting decrease in EWMA,
- * a constant smoothing factor between 0 and 1. The value
- * is used below for measuring average latency.
- */
- const float alpha = 0.2;
+ uint64_t now, latency;
+ static uint64_t prev_latency;
+ static bool first_sample = true;
now = rte_rdtsc();
@@ -170,32 +175,45 @@ calc_latency(uint16_t pid __rte_unused,
latency = now - *timestamp_dynfield(pkts[i]);
- /*
- * The jitter is calculated as statistical mean of interpacket
- * delay variation. The "jitter estimate" is computed by taking
- * the absolute values of the ipdv sequence and applying an
- * exponential filter with parameter 1/16 to generate the
- * estimate. i.e J=J+(|D(i-1,i)|-J)/16. Where J is jitter,
- * D(i-1,i) is difference in latency of two consecutive packets
- * i-1 and i.
- * Reference: Calculated as per RFC 5481, sec 4.1,
- * RFC 3393 sec 4.5, RFC 1889 sec.
- */
- glob_stats->jitter += (fabsf(prev_latency - latency)
- - glob_stats->jitter)/16;
- if (glob_stats->min_latency == 0)
- glob_stats->min_latency = latency;
- else if (latency < glob_stats->min_latency)
+ if (unlikely(first_sample)) {
+ first_sample = false;
+
glob_stats->min_latency = latency;
- else if (latency > glob_stats->max_latency)
glob_stats->max_latency = latency;
- /*
- * The average latency is measured using exponential moving
- * average, i.e. using EWMA
- * https://en.wikipedia.org/wiki/Moving_average
- */
- glob_stats->avg_latency +=
- alpha * (latency - glob_stats->avg_latency);
+ glob_stats->avg_latency = latency * 4;
+ /* start ad if previous sample had 0 latency */
+ glob_stats->jitter = latency / LATENCY_JITTER_SCALE;
+ } else {
+ /*
+ * The jitter is calculated as statistical mean of interpacket
+ * delay variation. The "jitter estimate" is computed by taking
+ * the absolute values of the ipdv sequence and applying an
+ * exponential filter with parameter 1/16 to generate the
+ * estimate. i.e J=J+(|D(i-1,i)|-J)/16. Where J is jitter,
+ * D(i-1,i) is difference in latency of two consecutive packets
+ * i-1 and i. Jitter is scaled by 16.
+ * Reference: Calculated as per RFC 5481, sec 4.1,
+ * RFC 3393 sec 4.5, RFC 1889 sec.
+ */
+ long long delta = prev_latency - latency;
+ glob_stats->jitter += llabs(delta)
+ - glob_stats->jitter / LATENCY_JITTER_SCALE;
+
+ if (latency < glob_stats->min_latency)
+ glob_stats->min_latency = latency;
+ if (latency > glob_stats->max_latency)
+ glob_stats->max_latency = latency;
+ /*
+ * The average latency is measured using exponential moving
+ * average, i.e. using EWMA
+ * https://en.wikipedia.org/wiki/Moving_average
+ *
+ * Alpha is .25, avg_latency is scaled by 4.
+ */
+ glob_stats->avg_latency += latency
+ - glob_stats->avg_latency / LATENCY_AVG_SCALE;
+ }
+
prev_latency = latency;
}
rte_spinlock_unlock(&glob_stats->lock);
--
2.43.0
next prev parent reply other threads:[~2024-04-22 15:24 UTC|newest]
Thread overview: 32+ messages / expand[flat|nested] mbox.gz Atom feed top
2024-04-08 19:50 [RFC] latencystats: performance overhaul Stephen Hemminger
2024-04-08 21:26 ` [PATCH v2] " Stephen Hemminger
2024-04-19 17:28 ` [PATCH v4 0/6] latencystats: cleanup Stephen Hemminger
2024-04-19 17:28 ` [PATCH v4 1/6] latencystats: replace use of VLA Stephen Hemminger
2024-04-19 17:28 ` [PATCH v4 2/6] latencystats: handle fractional cycles per ns Stephen Hemminger
2024-04-19 17:28 ` [PATCH v4 3/6] latencystats: do not use floating point Stephen Hemminger
2024-04-19 18:49 ` Morten Brørup
2024-04-19 22:45 ` Stephen Hemminger
2024-04-20 7:31 ` Morten Brørup
2024-04-19 17:28 ` [PATCH v4 4/6] latencystats: fix log messages Stephen Hemminger
2024-04-19 17:28 ` [PATCH v4 5/6] latencystats: update include files Stephen Hemminger
2024-04-19 17:28 ` [PATCH v4 6/6] latencystats: fix for pedantic warnings Stephen Hemminger
2024-04-22 15:21 ` [PATCH v5 0/9] latencystats: improve algorithms and tests Stephen Hemminger
2024-04-22 15:21 ` [PATCH v5 1/9] latencystats: replace use of VLA Stephen Hemminger
2024-04-22 15:21 ` [PATCH v5 2/9] latencystats: handle fractional cycles per ns Stephen Hemminger
2024-04-22 15:21 ` Stephen Hemminger [this message]
2024-04-22 15:21 ` [PATCH v5 4/9] latencystats: fix log messages Stephen Hemminger
2024-04-22 15:21 ` [PATCH v5 5/9] latencystats: update include files Stephen Hemminger
2024-04-22 15:21 ` [PATCH v5 6/9] latencystats: enforce that unused callback function is NULL Stephen Hemminger
2024-04-22 15:21 ` [PATCH v5 7/9] latencystats: add metric for number of samples Stephen Hemminger
2024-04-22 15:21 ` [PATCH v5 8/9] test: use initialization in latencystats test Stephen Hemminger
2024-04-22 15:21 ` [PATCH v5 9/9] test: add more latencystats tests Stephen Hemminger
2024-05-29 22:54 ` [PATCH v6 0/8] latencystats: improvements to algorithm and test Stephen Hemminger
2024-05-29 22:54 ` [PATCH v6 1/8] latencystats: replace use of VLA Stephen Hemminger
2024-05-29 22:54 ` [PATCH v6 2/8] latencystats: handle fractional cycles per ns Stephen Hemminger
2024-05-29 22:54 ` [PATCH v6 3/8] latencystats: do not use floating point Stephen Hemminger
2024-05-29 22:54 ` [PATCH v6 4/8] latencystats: fix log messages Stephen Hemminger
2024-05-29 22:54 ` [PATCH v6 5/8] latencystats: update include files Stephen Hemminger
2024-05-29 22:54 ` [PATCH v6 6/8] latencystats: enforce that unused callback function is NULL Stephen Hemminger
2024-05-29 22:54 ` [PATCH v6 7/8] latencystats: add metric for number of samples Stephen Hemminger
2024-05-29 22:54 ` [PATCH v6 8/8] test: update to latencystats tests Stephen Hemminger
2024-07-04 15:21 ` [PATCH v6 0/8] latencystats: improvements to algorithm and test David Marchand
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20240422152336.147553-4-stephen@networkplumber.org \
--to=stephen@networkplumber.org \
--cc=dev@dpdk.org \
--cc=mb@smartsharesystems.com \
--cc=reshma.pattan@intel.com \
--cc=roretzla@linux.microsoft.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).