From patchwork Wed May 29 22:54:40 2024 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 8bit X-Patchwork-Submitter: Stephen Hemminger X-Patchwork-Id: 140403 X-Patchwork-Delegate: thomas@monjalon.net Return-Path: X-Original-To: patchwork@inbox.dpdk.org Delivered-To: patchwork@inbox.dpdk.org Received: from mails.dpdk.org (mails.dpdk.org [217.70.189.124]) by inbox.dpdk.org (Postfix) with ESMTP id DFBBE44107; Thu, 30 May 2024 00:57:57 +0200 (CEST) Received: from mails.dpdk.org (localhost [127.0.0.1]) by mails.dpdk.org (Postfix) with ESMTP id 64507406FF; Thu, 30 May 2024 00:57:40 +0200 (CEST) Received: from mail-pg1-f176.google.com (mail-pg1-f176.google.com [209.85.215.176]) by mails.dpdk.org (Postfix) with ESMTP id C9DA0402DF for ; Thu, 30 May 2024 00:57:37 +0200 (CEST) Received: by mail-pg1-f176.google.com with SMTP id 41be03b00d2f7-6bce380eb96so207251a12.0 for ; Wed, 29 May 2024 15:57:37 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=networkplumber-org.20230601.gappssmtp.com; s=20230601; t=1717023457; x=1717628257; darn=dpdk.org; h=content-transfer-encoding:mime-version:references:in-reply-to :message-id:date:subject:cc:to:from:from:to:cc:subject:date :message-id:reply-to; bh=/dspqKZ7Gwydk7sWgMWJ84/aRMy8m2p+7G7/8P0RBFo=; b=TeI1up+gYjjKd4/e+F78OIxI+fFZnOT4hgcQx3cKKwtkkmbgbhg3jqDraSVvnGVVjS juCM7bToJzTY4+B0zgUavIQ1o13cQU4l7vPT9wA4bl0iXiaCUFICjRkWQ5J+mGCZyYVj ZO7HfhN4yB1yFyU96C744fe4jOexm0U2pdmc7lq51JTWNEEoAgRn+iXAwDJAyYkxWhFw ycZuXjydLBoaD+bMfU+iLVoBKOEiUIQ7XyoEPP0GM0XVNOemx1lhRwUSQ/FdF00q7RMK A26+9+YCJEKxtG0m7927Blbdda3ESx13CoJ3i+x9hfx8QsmLdkxfcUJCR5FMQGxxc3Ag TjMw== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20230601; t=1717023457; x=1717628257; h=content-transfer-encoding:mime-version:references:in-reply-to :message-id:date:subject:cc:to:from:x-gm-message-state:from:to:cc :subject:date:message-id:reply-to; bh=/dspqKZ7Gwydk7sWgMWJ84/aRMy8m2p+7G7/8P0RBFo=; b=cSV0bMCAOikxo5t/+QO3N9d7qu/TyMfKkwaDnLhFyHungOPovDoWACwxOFva1+G79E eyDD9zxAHd13f5B93E9nY+SLroQl/yzblz64E+PhaopVWBd0T+me1DRUxsyqxMouO1iL PBuHbRPOaP4L2EFqjTUsRuCh/PcDVxZIqj1HJS4m8SeQDhjxgP/QhTAkuJgyNRx2K0Dj PWQpX5UQlOrIlqoaNU+9N0iL+YWt1hKKv6g5j/5f7cu268tJm2u7rQ2dVicU1u8U0k9U l1OyrfCRMOTIpX0ySpdLqt3MApmsM79wr1B+zzmGbFUlZkrIbftYTOwVgXTs3uQKPmdj rHvQ== X-Gm-Message-State: AOJu0YxS1CAw7Mrr96qH566IYDXLUqlblHDfNf46ArdduT9auFIVrANu 6HnEkIBAYhOplXi6lbfnp6ICVNXBfa48uco/4ebdMEbwnDQRddjkee9i+qcm7ZkhVSltNFQbDd5 N X-Google-Smtp-Source: AGHT+IG8ay/QLI+iRR0R/45ObSF+CricOfFYPRrV/t4xDBlp9n60pYw/26sYjkRdFCQU4JpZENePzQ== X-Received: by 2002:a17:90b:17c4:b0:2bf:9981:e0bc with SMTP id 98e67ed59e1d1-2c1abbfb17bmr472846a91.27.1717023456843; Wed, 29 May 2024 15:57:36 -0700 (PDT) Received: from hermes.lan (204-195-96-226.wavecable.com. [204.195.96.226]) by smtp.gmail.com with ESMTPSA id 98e67ed59e1d1-2c1a77baebcsm351762a91.53.2024.05.29.15.57.36 (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Wed, 29 May 2024 15:57:36 -0700 (PDT) From: Stephen Hemminger To: dev@dpdk.org Cc: Stephen Hemminger , Tyler Retzlaff , =?utf-8?q?Morten_Br=C3=B8rup?= , Reshma Pattan Subject: [PATCH v6 3/8] latencystats: do not use floating point Date: Wed, 29 May 2024 15:54:40 -0700 Message-ID: <20240529225732.442539-4-stephen@networkplumber.org> X-Mailer: git-send-email 2.43.0 In-Reply-To: <20240529225732.442539-1-stephen@networkplumber.org> References: <20240408195036.182545-1-stephen@networkplumber.org> <20240529225732.442539-1-stephen@networkplumber.org> MIME-Version: 1.0 X-BeenThere: dev@dpdk.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: DPDK patches and discussions List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces@dpdk.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 Acked-by: Tyler Retzlaff Acked-by: Morten Brørup --- 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 c897df03c6..2140ab87c4 100644 --- a/lib/latencystats/rte_latencystats.c +++ b/lib/latencystats/rte_latencystats.c @@ -4,6 +4,7 @@ #include +#include #include #include #include @@ -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 = 0.2f; + 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);