DAOS-14105 object: latency metrics for collective punch

For locating the performance bottle neck.

Signed-off-by: Fan Yong <fan.yong@intel.com>

src/include/daos_srv/dtx_srv.h

@@ -152,7 +152,8 @@ struct dtx_coll_entry {
 	uint8_t				*dce_hints;
 	uint8_t				*dce_bitmap;
 	uint32_t			 dce_hint_sz;
-	uint32_t			 dce_bitmap_sz;
+	uint16_t			 dce_bitmap_sz;
+	uint16_t			 dce_tgt_nr;
 };
 
 struct dtx_leader_handle;

src/object/obj_rpc.h

@@ -185,6 +185,8 @@ enum obj_rpc_flags {
 	ORF_REBUILDING_IO	= (1 << 23),
 	/* 'sgls' is NULL, for update sub-request of CPD RPC. */
 	ORF_EMPTY_SGL		= (1 << 24),
+	/* Internal transaction for punch object. */
+	ORF_INTERNAL_PUNCH	= (1 << 25),
 };
 
 /* common for update/fetch */

src/object/obj_tx.c

@@ -84,6 +84,7 @@ struct dc_tx {
 				 tx_retry:1, /** Retry the commit RPC. */
 				 tx_set_resend:1, /** Set 'resend' flag. */
 				 tx_for_convert:1,
+				 tx_internal_punch:1,
 				 tx_has_cond:1,
 				 tx_renew:1,
 				 tx_closed:1,
@@ -1918,6 +1919,9 @@ dc_tx_commit_prepare(struct dc_tx *tx, tse_task_t *task)
 				if (rc != 0)
 					goto out;
 			}
+
+			if (tx->tx_for_convert)
+				tx->tx_internal_punch = 1;
 		} else {
 			grp_idx = obj_dkey2grpidx(obj, dcsr->dcsr_dkey_hash,
 						  tx->tx_pm_ver);
@@ -2311,6 +2315,8 @@ dc_tx_commit_trigger(tse_task_t *task, struct dc_tx *tx, daos_tx_commit_t *args)
 	tx->tx_renew = 0;
 	if (tx->tx_reintegrating)
 		oci->oci_flags |= ORF_REINTEGRATING_IO;
+	if (tx->tx_internal_punch)
+		oci->oci_flags |= ORF_INTERNAL_PUNCH;
 
 	oci->oci_sub_heads.ca_arrays = &tx->tx_head;
 	oci->oci_sub_heads.ca_count = 1;

src/object/srv_internal.h

@@ -160,6 +160,16 @@ struct obj_tls {
 
 	struct d_tm_node_t	*ot_update_bio_lat[NR_LATENCY_BUCKETS];
 	struct d_tm_node_t	*ot_fetch_bio_lat[NR_LATENCY_BUCKETS];
+
+	struct d_tm_node_t	*ot_coll_punch_map_lat[NR_LATENCY_BUCKETS];
+	struct d_tm_node_t	*ot_coll_punch_layout_lat[NR_LATENCY_BUCKETS];
+	struct d_tm_node_t	*ot_coll_punch_parse_lat[NR_LATENCY_BUCKETS];
+	struct d_tm_node_t	*ot_coll_punch_lexec_lat[NR_LATENCY_BUCKETS];
+	struct d_tm_node_t	*ot_coll_punch_texec_lat[NR_LATENCY_BUCKETS];
+
+	struct d_tm_node_t	*ot_cpd_punch_lexec_lat[NR_LATENCY_BUCKETS];
+	struct d_tm_node_t	*ot_cpd_punch_texec_lat[NR_LATENCY_BUCKETS];
+	struct d_tm_node_t	*ot_cpd_others_lat[NR_LATENCY_BUCKETS];
 };
 
 static inline struct obj_tls *
@@ -186,55 +196,28 @@ lat_bucket(uint64_t size)
 	return 56 - nr;
 }
 
+static inline int
+lat_log_bucket(uint64_t size)
+{
+	int	i;
+
+	for (i = 0; size > 0; i++)
+		size >>= 1;
+
+	return i >= NR_LATENCY_BUCKETS ? NR_LATENCY_BUCKETS - 1 : i;
+}
+
 enum latency_type {
 	BULK_LATENCY,
 	BIO_LATENCY,
 	VOS_LATENCY,
+	FIND_MAP,
+	GEN_LAYOUT,
+	COLL_PARSE,
+	LOCAL_EXEC,
+	TOTAL_EXEC,
 };
 
-static inline void
-obj_update_latency(uint32_t opc, uint32_t type, uint64_t latency, uint64_t io_size)
-{
-	struct obj_tls		*tls = obj_tls_get();
-	struct d_tm_node_t	*lat;
-
-	latency >>= 10; /* convert to micro seconds */
-
-	if (opc == DAOS_OBJ_RPC_FETCH) {
-		switch (type) {
-		case BULK_LATENCY:
-			lat = tls->ot_fetch_bulk_lat[lat_bucket(io_size)];
-			break;
-		case BIO_LATENCY:
-			lat = tls->ot_fetch_bio_lat[lat_bucket(io_size)];
-			break;
-		case VOS_LATENCY:
-			lat = tls->ot_fetch_vos_lat[lat_bucket(io_size)];
-			break;
-		default:
-			D_ASSERT(0);
-		}
-	} else if (opc == DAOS_OBJ_RPC_UPDATE || opc == DAOS_OBJ_RPC_TGT_UPDATE) {
-		switch (type) {
-		case BULK_LATENCY:
-			lat = tls->ot_update_bulk_lat[lat_bucket(io_size)];
-			break;
-		case BIO_LATENCY:
-			lat = tls->ot_update_bio_lat[lat_bucket(io_size)];
-			break;
-		case VOS_LATENCY:
-			lat = tls->ot_update_vos_lat[lat_bucket(io_size)];
-			break;
-		default:
-			D_ASSERT(0);
-		}
-	} else {
-		/* Ignore other ops for the moment */
-		return;
-	}
-	d_tm_set_gauge(lat, latency);
-}
-
 struct ds_obj_exec_arg {
 	crt_rpc_t		*rpc;
 	struct obj_io_context	*ioc;

src/object/srv_mod.c

@@ -125,10 +125,6 @@ obj_tls_init(int tags, int xs_id, int tgt_id)
 
 	D_INIT_LIST_HEAD(&tls->ot_pool_list);
 
-	if (tgt_id < 0)
-		/** skip sensor setup on system xstreams */
-		return tls;
-
 	/** register different per-opcode sensors */
 	for (opc = 0; opc < OBJ_PROTO_CLI_COUNT; opc++) {
 		/** Start with number of active requests, of type gauge */
@@ -140,9 +136,8 @@ obj_tls_init(int tags, int xs_id, int tgt_id)
 			D_WARN("Failed to create active counter: "DF_RC"\n",
 			       DP_RC(rc));
 
-		if (opc == DAOS_OBJ_RPC_UPDATE ||
-		    opc == DAOS_OBJ_RPC_TGT_UPDATE ||
-		    opc == DAOS_OBJ_RPC_FETCH)
+		if (opc == DAOS_OBJ_RPC_UPDATE || opc == DAOS_OBJ_RPC_TGT_UPDATE ||
+		    opc == DAOS_OBJ_RPC_FETCH || opc == DAOS_OBJ_RPC_COLL_PUNCH)
 			/** See below, latency reported per size for those */
 			continue;
 
@@ -184,6 +179,24 @@ obj_tls_init(int tags, int xs_id, int tgt_id)
 	obj_latency_tm_init(DAOS_OBJ_RPC_FETCH, tgt_id, tls->ot_fetch_bio_lat,
 			    "bio_fetch", "BIO fetch processing time");
 
+	obj_latency_tm_init(DAOS_OBJ_RPC_COLL_PUNCH, tgt_id, tls->ot_coll_punch_map_lat,
+			    "coll_punch_map", "Find pool map processing time");
+	obj_latency_tm_init(DAOS_OBJ_RPC_COLL_PUNCH, tgt_id, tls->ot_coll_punch_layout_lat,
+			    "coll_punch_layout", "Generate object layout processing time");
+	obj_latency_tm_init(DAOS_OBJ_RPC_COLL_PUNCH, tgt_id, tls->ot_coll_punch_parse_lat,
+			    "coll_punch_parse", "Parse layout processing time");
+	obj_latency_tm_init(DAOS_OBJ_RPC_COLL_PUNCH, tgt_id, tls->ot_coll_punch_lexec_lat,
+			    "coll_punch_lexec", "Local collective exec processing time");
+	obj_latency_tm_init(DAOS_OBJ_RPC_COLL_PUNCH, tgt_id, tls->ot_coll_punch_texec_lat,
+			    "coll_punch_texec", "Total collective exec processing time");
+
+	obj_latency_tm_init(DAOS_OBJ_RPC_CPD, tgt_id, tls->ot_cpd_punch_lexec_lat,
+			    "cpd_punch_lexec", "Local cpd punch processing time");
+	obj_latency_tm_init(DAOS_OBJ_RPC_CPD, tgt_id, tls->ot_cpd_punch_texec_lat,
+			    "cpd_punch_texec", "Total cpd punch processing time");
+	obj_latency_tm_init(DAOS_OBJ_RPC_CPD, tgt_id, tls->ot_cpd_others_lat,
+			    "cpd_others", "Other CPD operation processing time");
+
 	return tls;
 }