bgpd: improve labelpool performance at scale

    - double the size of each new chunk request from zebra
    - use bitfields to track label allocations in a chunk
    - When allocating:
        - skip chunks with no free labels
        - search biggest chunks first
        - start search in chunk where last search ended
    - Improve API documentation in comments (bgp_lp_get() and callback)
    - Tweak formatting of "show bgp labelpool chunks"
    - Add test features (compiled conditionally on BGP_LABELPOOL_ENABLE_TESTS)

Signed-off-by: G. Paul Ziemba <paulz@labn.net>

6 files changed, 693 insertions, 29 deletions

diff --git a/bgpd/bgp_labelpool.c b/bgpd/bgp_labelpool.c
index fa1dcf33e..c227a5e41 100644
--- a/bgpd/bgp_labelpool.c
+++ b/bgpd/bgp_labelpool.c
@@ -37,18 +37,41 @@
 #include "bgpd/bgp_errors.h"
 #include "bgpd/bgp_route.h"
 
+#define BGP_LABELPOOL_ENABLE_TESTS 0
+
+#ifndef VTYSH_EXTRACT_PL
+#include "bgpd/bgp_labelpool_clippy.c"
+#endif
+
+
 /*
  * Definitions and external declarations.
  */
 extern struct zclient *zclient;
 
+#if BGP_LABELPOOL_ENABLE_TESTS
+static void lptest_init(void);
+static void lptest_finish(void);
+#endif
+
 /*
  * Remember where pool data are kept
  */
 static struct labelpool *lp;
 
-/* request this many labels at a time from zebra */
-#define LP_CHUNK_SIZE	50
+/*
+ * Number of labels requested at a time from the zebra label manager.
+ * We start small but double the request size each time up to a
+ * maximum size.
+ *
+ * The label space is 20 bits which is shared with other FRR processes
+ * on this host, so to avoid greedily requesting a mostly wasted chunk,
+ * we limit the chunk size to 1/16 of the label space (that's the -4 bits
+ * in the definition below). This limit slightly increases our cost of
+ * finding free labels in our allocated chunks.
+ */
+#define LP_CHUNK_SIZE_MIN 128
+#define LP_CHUNK_SIZE_MAX (1 << (20 - 4))
 
 DEFINE_MTYPE_STATIC(BGPD, BGP_LABEL_CHUNK, "BGP Label Chunk");
 DEFINE_MTYPE_STATIC(BGPD, BGP_LABEL_FIFO, "BGP Label FIFO item");
@@ -58,6 +81,9 @@ DEFINE_MTYPE_STATIC(BGPD, BGP_LABEL_CBQ, "BGP Dynamic Label Callback");
 struct lp_chunk {
 	uint32_t	first;
 	uint32_t	last;
+	uint32_t nfree;		     /* un-allocated count */
+	uint32_t idx_last_allocated; /* start looking here */
+	bitfield_t allocated_map;
 };
 
 /*
@@ -160,6 +186,9 @@ static void lp_lcb_free(void *goner)
 
 static void lp_chunk_free(void *goner)
 {
+	struct lp_chunk *chunk = (struct lp_chunk *)goner;
+
+	bf_free(chunk->allocated_map);
 	XFREE(MTYPE_BGP_LABEL_CHUNK, goner);
 }
 
@@ -180,6 +209,12 @@ void bgp_lp_init(struct thread_master *master, struct labelpool *pool)
 	lp->callback_q->spec.workfunc = lp_cbq_docallback;
 	lp->callback_q->spec.del_item_data = lp_cbq_item_free;
 	lp->callback_q->spec.max_retries = 0;
+
+	lp->next_chunksize = LP_CHUNK_SIZE_MIN;
+
+#if BGP_LABELPOOL_ENABLE_TESTS
+	lptest_init();
+#endif
 }
 
 /* check if a label callback was for a BGP LU node, and if so, unlock it */
@@ -201,6 +236,9 @@ void bgp_lp_finish(void)
 	struct lp_fifo *lf;
 	struct work_queue_item *item, *titem;
 
+#if BGP_LABELPOOL_ENABLE_TESTS
+	lptest_finish();
+#endif
 	if (!lp)
 		return;
 
@@ -240,25 +278,53 @@ static mpls_label_t get_label_from_pool(void *labelid)
 
 	/*
 	 * Find a free label
-	 * Linear search is not efficient but should be executed infrequently.
 	 */
 	for (ALL_LIST_ELEMENTS_RO(lp->chunks, node, chunk)) {
 		uintptr_t lbl;
+		unsigned int index;
 
 		if (debug)
 			zlog_debug("%s: chunk first=%u last=%u",
 				__func__, chunk->first, chunk->last);
 
-		for (lbl = chunk->first; lbl <= chunk->last; ++lbl) {
-			/* labelid is key to all-request "ledger" list */
-			if (!skiplist_insert(lp->inuse, (void *)lbl, labelid)) {
-				/*
-				 * Success
-				 */
-				return lbl;
-			}
+		/*
+		 * don't look in chunks with no available labels
+		 */
+		if (!chunk->nfree)
+			continue;
+
+		/*
+		 * roll through bitfield starting where we stopped
+		 * last time
+		 */
+		index = bf_find_next_clear_bit_wrap(
+			&chunk->allocated_map, chunk->idx_last_allocated + 1,
+			0);
+
+		/*
+		 * since chunk->nfree is non-zero, we should always get
+		 * a valid index
+		 */
+		assert(index != WORD_MAX);
+
+		lbl = chunk->first + index;
+		if (skiplist_insert(lp->inuse, (void *)lbl, labelid)) {
+			/* something is very wrong */
+			zlog_err("%s: unable to insert inuse label %u (id %p)",
+				 __func__, (uint32_t)lbl, labelid);
+			return MPLS_LABEL_NONE;
 		}
+
+		/*
+		 * Success
+		 */
+		bf_set_bit(chunk->allocated_map, index);
+		chunk->idx_last_allocated = index;
+		chunk->nfree -= 1;
+
+		return lbl;
 	}
+
 	return MPLS_LABEL_NONE;
 }
 
@@ -299,10 +365,14 @@ static struct lp_lcb *lcb_alloc(
  * When connection to zebra is reestablished, previous label assignments
  * will be invalidated (via callbacks having the "allocated" parameter unset)
  * and new labels will be automatically reassigned by this labelpool module
- * (that is, a requestor does not need to call lp_get() again if it is
+ * (that is, a requestor does not need to call bgp_lp_get() again if it is
  * notified via callback that its label has been lost: it will eventually
  * get another callback with a new label assignment).
  *
+ * The callback function should return 0 to accept the allocation
+ * and non-zero to refuse it. The callback function return value is
+ * ignored for invalidations (i.e., when the "allocated" parameter is false)
+ *
  * Prior requests for a given labelid are detected so that requests and
  * assignments are not duplicated.
  */
@@ -392,10 +462,13 @@ void bgp_lp_get(
 	if (lp_fifo_count(&lp->requests) > lp->pending_count) {
 		if (!zclient || zclient->sock < 0)
 			return;
-		if (zclient_send_get_label_chunk(zclient, 0, LP_CHUNK_SIZE,
-						 MPLS_LABEL_BASE_ANY)
-		    != ZCLIENT_SEND_FAILURE)
-			lp->pending_count += LP_CHUNK_SIZE;
+		if (zclient_send_get_label_chunk(zclient, 0, lp->next_chunksize,
+						 MPLS_LABEL_BASE_ANY) !=
+		    ZCLIENT_SEND_FAILURE) {
+			lp->pending_count += lp->next_chunksize;
+			if ((lp->next_chunksize << 1) <= LP_CHUNK_SIZE_MAX)
+				lp->next_chunksize <<= 1;
+		}
 	}
 }
 
@@ -408,13 +481,36 @@ void bgp_lp_release(
 
 	if (!skiplist_search(lp->ledger, labelid, (void **)&lcb)) {
 		if (label == lcb->label && type == lcb->type) {
+			struct listnode *node;
+			struct lp_chunk *chunk;
 			uintptr_t lbl = label;
+			bool deallocated = false;
 
 			/* no longer in use */
 			skiplist_delete(lp->inuse, (void *)lbl, NULL);
 
 			/* no longer requested */
 			skiplist_delete(lp->ledger, labelid, NULL);
+
+			/*
+			 * Find the chunk this label belongs to and
+			 * deallocate the label
+			 */
+			for (ALL_LIST_ELEMENTS_RO(lp->chunks, node, chunk)) {
+				uint32_t index;
+
+				if ((label < chunk->first) ||
+				    (label > chunk->last))
+					continue;
+
+				index = label - chunk->first;
+				assert(bf_test_index(chunk->allocated_map,
+						     index));
+				bf_release_index(chunk->allocated_map, index);
+				chunk->nfree += 1;
+				deallocated = true;
+			}
+			assert(deallocated);
 		}
 	}
 }
@@ -427,6 +523,7 @@ void bgp_lp_event_chunk(uint8_t keep, uint32_t first, uint32_t last)
 	struct lp_chunk *chunk;
 	int debug = BGP_DEBUG(labelpool, LABELPOOL);
 	struct lp_fifo *lf;
+	uint32_t labelcount;
 
 	if (last < first) {
 		flog_err(EC_BGP_LABEL,
@@ -437,19 +534,27 @@ void bgp_lp_event_chunk(uint8_t keep, uint32_t first, uint32_t last)
 
 	chunk = XCALLOC(MTYPE_BGP_LABEL_CHUNK, sizeof(struct lp_chunk));
 
+	labelcount = last - first + 1;
+
 	chunk->first = first;
 	chunk->last = last;
+	chunk->nfree = labelcount;
+	bf_init(chunk->allocated_map, labelcount);
 
-	listnode_add(lp->chunks, chunk);
+	/*
+	 * Optimize for allocation by adding the new (presumably larger)
+	 * chunk at the head of the list so it is examined first.
+	 */
+	listnode_add_head(lp->chunks, chunk);
 
-	lp->pending_count -= (last - first + 1);
+	lp->pending_count -= labelcount;
 
 	if (debug) {
 		zlog_debug("%s: %zu pending requests", __func__,
 			lp_fifo_count(&lp->requests));
 	}
 
-	while ((lf = lp_fifo_first(&lp->requests))) {
+	while (labelcount && (lf = lp_fifo_first(&lp->requests))) {
 
 		struct lp_lcb *lcb;
 		void *labelid = lf->lcb.labelid;
@@ -495,6 +600,8 @@ void bgp_lp_event_chunk(uint8_t keep, uint32_t first, uint32_t last)
 			break;
 		}
 
+		labelcount -= 1;
+
 		/*
 		 * we filled the request from local pool.
 		 * Enqueue response work item with new label.
@@ -535,8 +642,8 @@ void bgp_lp_event_zebra_down(void)
  */
 void bgp_lp_event_zebra_up(void)
 {
-	int labels_needed;
-	int chunks_needed;
+	unsigned int labels_needed;
+	unsigned int chunks_needed;
 	void *labelid;
 	struct lp_lcb *lcb;
 	int lm_init_ok;
@@ -548,9 +655,16 @@ void bgp_lp_event_zebra_up(void)
 	labels_needed = lp_fifo_count(&lp->requests) +
 		skiplist_count(lp->inuse);
 
+	if (labels_needed > lp->next_chunksize) {
+		while ((lp->next_chunksize < labels_needed) &&
+		       (lp->next_chunksize << 1 <= LP_CHUNK_SIZE_MAX))
+
+			lp->next_chunksize <<= 1;
+	}
+
 	/* round up */
-	chunks_needed = (labels_needed / LP_CHUNK_SIZE) + 1;
-	labels_needed = chunks_needed * LP_CHUNK_SIZE;
+	chunks_needed = (labels_needed / lp->next_chunksize) + 1;
+	labels_needed = chunks_needed * lp->next_chunksize;
 
 	lm_init_ok = lm_label_manager_connect(zclient, 1) == 0;
 
@@ -937,25 +1051,496 @@ DEFUN(show_bgp_labelpool_chunks, show_bgp_labelpool_chunks_cmd,
 		}
 		json = json_object_new_array();
 	} else {
-		vty_out(vty, "First    Last\n");
-		vty_out(vty, "--------------\n");
+		vty_out(vty, "%10s %10s %10s %10s\n", "First", "Last", "Size",
+			"nfree");
+		vty_out(vty, "-------------------------------------------\n");
 	}
 
 	for (ALL_LIST_ELEMENTS_RO(lp->chunks, node, chunk)) {
+		uint32_t size;
+
+		size = chunk->last - chunk->first + 1;
+
 		if (uj) {
 			json_elem = json_object_new_object();
 			json_object_array_add(json, json_elem);
 			json_object_int_add(json_elem, "first", chunk->first);
 			json_object_int_add(json_elem, "last", chunk->last);
+			json_object_int_add(json_elem, "size", size);
+			json_object_int_add(json_elem, "numberFree",
+					    chunk->nfree);
 		} else
-			vty_out(vty, "%-10u %-10u\n", chunk->first,
-				chunk->last);
+			vty_out(vty, "%10u %10u %10u %10u\n", chunk->first,
+				chunk->last, size, chunk->nfree);
 	}
 	if (uj)
 		vty_json(vty, json);
 	return CMD_SUCCESS;
 }
 
+#if BGP_LABELPOOL_ENABLE_TESTS
+/*------------------------------------------------------------------------
+ *			Testing code start
+ *------------------------------------------------------------------------*/
+
+DEFINE_MTYPE_STATIC(BGPD, LABELPOOL_TEST, "Label pool test");
+
+#define LPT_STAT_INSERT_FAIL 0
+#define LPT_STAT_DELETE_FAIL 1
+#define LPT_STAT_ALLOCATED 2
+#define LPT_STAT_DEALLOCATED 3
+#define LPT_STAT_MAX 4
+
+const char *lpt_counter_names[] = {
+	"sl insert failures",
+	"sl delete failures",
+	"labels allocated",
+	"labels deallocated",
+};
+
+static uint8_t lpt_generation;
+static bool lpt_inprogress;
+static struct skiplist *lp_tests;
+static unsigned int lpt_test_cb_tcb_lookup_fails;
+static unsigned int lpt_release_tcb_lookup_fails;
+static unsigned int lpt_test_event_tcb_lookup_fails;
+static unsigned int lpt_stop_tcb_lookup_fails;
+
+struct lp_test {
+	uint8_t generation;
+	unsigned int request_maximum;
+	unsigned int request_blocksize;
+	uintptr_t request_count; /* match type of labelid */
+	int label_type;
+	struct skiplist *labels;
+	struct timeval starttime;
+	struct skiplist *timestamps_alloc;
+	struct skiplist *timestamps_dealloc;
+	struct thread *event_thread;
+	unsigned int counter[LPT_STAT_MAX];
+};
+
+/* test parameters */
+#define LPT_MAX_COUNT 500000  /* get this many labels in all */
+#define LPT_BLKSIZE 10000     /* this many at a time, then yield */
+#define LPT_TS_INTERVAL 10000 /* timestamp every this many labels */
+
+
+static int test_cb(mpls_label_t label, void *labelid, bool allocated)
+{
+	uintptr_t generation;
+	struct lp_test *tcb;
+
+	generation = ((uintptr_t)labelid >> 24) & 0xff;
+
+	if (skiplist_search(lp_tests, (void *)generation, (void **)&tcb)) {
+
+		/* couldn't find current test in progress */
+		++lpt_test_cb_tcb_lookup_fails;
+		return -1; /* reject allocation */
+	}
+
+	if (allocated) {
+		++tcb->counter[LPT_STAT_ALLOCATED];
+		if (!(tcb->counter[LPT_STAT_ALLOCATED] % LPT_TS_INTERVAL)) {
+			uintptr_t time_ms;
+
+			time_ms = monotime_since(&tcb->starttime, NULL) / 1000;
+			skiplist_insert(tcb->timestamps_alloc,
+					(void *)(uintptr_t)tcb
+						->counter[LPT_STAT_ALLOCATED],
+					(void *)time_ms);
+		}
+		if (skiplist_insert(tcb->labels, labelid,
+				    (void *)(uintptr_t)label)) {
+			++tcb->counter[LPT_STAT_INSERT_FAIL];
+			return -1;
+		}
+	} else {
+		++tcb->counter[LPT_STAT_DEALLOCATED];
+		if (!(tcb->counter[LPT_STAT_DEALLOCATED] % LPT_TS_INTERVAL)) {
+			uintptr_t time_ms;
+
+			time_ms = monotime_since(&tcb->starttime, NULL) / 1000;
+			skiplist_insert(tcb->timestamps_dealloc,
+					(void *)(uintptr_t)tcb
+						->counter[LPT_STAT_ALLOCATED],
+					(void *)time_ms);
+		}
+		if (skiplist_delete(tcb->labels, labelid, 0)) {
+			++tcb->counter[LPT_STAT_DELETE_FAIL];
+			return -1;
+		}
+	}
+	return 0;
+}
+
+static void labelpool_test_event_handler(struct thread *thread)
+{
+	struct lp_test *tcb;
+
+	if (skiplist_search(lp_tests, (void *)(uintptr_t)(lpt_generation),
+			    (void **)&tcb)) {
+
+		/* couldn't find current test in progress */
+		++lpt_test_event_tcb_lookup_fails;
+		return;
+	}
+
+	/*
+	 * request a bunch of labels
+	 */
+	for (unsigned int i = 0; (i < tcb->request_blocksize) &&
+				 (tcb->request_count < tcb->request_maximum);
+	     ++i) {
+
+		uintptr_t id;
+
+		++tcb->request_count;
+
+		/*
+		 * construct 32-bit id from request_count and generation
+		 */
+		id = ((uintptr_t)tcb->generation << 24) |
+		     (tcb->request_count & 0x00ffffff);
+		bgp_lp_get(LP_TYPE_VRF, (void *)id, test_cb);
+	}
+
+	if (tcb->request_count < tcb->request_maximum)
+		thread_add_event(bm->master, labelpool_test_event_handler, NULL,
+				 0, &tcb->event_thread);
+}
+
+static void lptest_stop(void)
+{
+	struct lp_test *tcb;
+
+	if (!lpt_inprogress)
+		return;
+
+	if (skiplist_search(lp_tests, (void *)(uintptr_t)(lpt_generation),
+			    (void **)&tcb)) {
+
+		/* couldn't find current test in progress */
+		++lpt_stop_tcb_lookup_fails;
+		return;
+	}
+
+	if (tcb->event_thread)
+		thread_cancel(&tcb->event_thread);
+
+	lpt_inprogress = false;
+}
+
+static int lptest_start(struct vty *vty)
+{
+	struct lp_test *tcb;
+
+	if (lpt_inprogress) {
+		vty_out(vty, "test already in progress\n");
+		return -1;
+	}
+
+	if (skiplist_count(lp_tests) >=
+	    (1 << (8 * sizeof(lpt_generation))) - 1) {
+		/*
+		 * Too many test runs
+		 */
+		vty_out(vty, "too many tests: clear first\n");
+		return -1;
+	}
+
+	/*
+	 * We pack the generation and request number into the labelid;
+	 * make sure they fit.
+	 */
+	unsigned int n1 = LPT_MAX_COUNT;
+	unsigned int sh = 0;
+	unsigned int label_bits;
+
+	label_bits = 8 * (sizeof(tcb->request_count) - sizeof(lpt_generation));
+
+	/* n1 should be same type as tcb->request_maximum */
+	assert(sizeof(n1) == sizeof(tcb->request_maximum));
+
+	while (n1 >>= 1)
+		++sh;
+	sh += 1; /* number of bits needed to hold LPT_MAX_COUNT */
+
+	if (sh > label_bits) {
+		vty_out(vty,
+			"Sorry, test iteration count too big on this platform (LPT_MAX_COUNT %u, need %u bits, but label_bits is only %u)\n",
+			LPT_MAX_COUNT, sh, label_bits);
+		return -1;
+	}
+
+	lpt_inprogress = true;
+	++lpt_generation;
+
+	tcb = XCALLOC(MTYPE_LABELPOOL_TEST, sizeof(*tcb));
+
+	tcb->generation = lpt_generation;
+	tcb->label_type = LP_TYPE_VRF;
+	tcb->request_maximum = LPT_MAX_COUNT;
+	tcb->request_blocksize = LPT_BLKSIZE;
+	tcb->labels = skiplist_new(0, NULL, NULL);
+	tcb->timestamps_alloc = skiplist_new(0, NULL, NULL);
+	tcb->timestamps_dealloc = skiplist_new(0, NULL, NULL);
+	thread_add_event(bm->master, labelpool_test_event_handler, NULL, 0,
+			 &tcb->event_thread);
+	monotime(&tcb->starttime);
+
+	skiplist_insert(lp_tests, (void *)(uintptr_t)tcb->generation, tcb);
+	return 0;
+}
+
+DEFPY(start_labelpool_perf_test, start_labelpool_perf_test_cmd,
+      "debug bgp lptest start",
+      DEBUG_STR BGP_STR
+      "label pool test\n"
+      "start\n")
+{
+	lptest_start(vty);
+	return CMD_SUCCESS;
+}
+
+static void lptest_print_stats(struct vty *vty, struct lp_test *tcb)
+{
+	unsigned int i;
+
+	vty_out(vty, "Global Lookup Failures in test_cb: %5u\n",
+		lpt_test_cb_tcb_lookup_fails);
+	vty_out(vty, "Global Lookup Failures in release: %5u\n",
+		lpt_release_tcb_lookup_fails);
+	vty_out(vty, "Global Lookup Failures in event:   %5u\n",
+		lpt_test_event_tcb_lookup_fails);
+	vty_out(vty, "Global Lookup Failures in stop:    %5u\n",
+		lpt_stop_tcb_lookup_fails);
+	vty_out(vty, "\n");
+
+	if (!tcb) {
+		if (skiplist_search(lp_tests, (void *)(uintptr_t)lpt_generation,
+				    (void **)&tcb)) {
+			vty_out(vty, "Error: can't find test %u\n",
+				lpt_generation);
+			return;
+		}
+	}
+
+	vty_out(vty, "Test Generation %u:\n", tcb->generation);
+
+	vty_out(vty, "Counter   Value\n");
+	for (i = 0; i < LPT_STAT_MAX; ++i) {
+		vty_out(vty, "%20s: %10u\n", lpt_counter_names[i],
+			tcb->counter[i]);
+	}
+	vty_out(vty, "\n");
+
+	if (tcb->timestamps_alloc) {
+		void *Key;
+		void *Value;
+		void *cursor;
+
+		float elapsed;
+
+		vty_out(vty, "%10s %10s\n", "Count", "Seconds");
+
+		cursor = NULL;
+		while (!skiplist_next(tcb->timestamps_alloc, &Key, &Value,
+				      &cursor)) {
+
+			elapsed = ((float)(uintptr_t)Value) / 1000;
+
+			vty_out(vty, "%10llu %10.3f\n",
+				(unsigned long long)(uintptr_t)Key, elapsed);
+		}
+		vty_out(vty, "\n");
+	}
+}
+
+DEFPY(show_labelpool_perf_test, show_labelpool_perf_test_cmd,
+      "debug bgp lptest show",
+      DEBUG_STR BGP_STR
+      "label pool test\n"
+      "show\n")
+{
+
+	if (lp_tests) {
+		void *Key;
+		void *Value;
+		void *cursor;
+
+		cursor = NULL;
+		while (!skiplist_next(lp_tests, &Key, &Value, &cursor)) {
+			lptest_print_stats(vty, (struct lp_test *)Value);
+		}
+	} else {
+		vty_out(vty, "no test results\n");
+	}
+	return CMD_SUCCESS;
+}
+
+DEFPY(stop_labelpool_perf_test, stop_labelpool_perf_test_cmd,
+      "debug bgp lptest stop",
+      DEBUG_STR BGP_STR
+      "label pool test\n"
+      "stop\n")
+{
+
+	if (lpt_inprogress) {
+		lptest_stop();
+		lptest_print_stats(vty, NULL);
+	} else {
+		vty_out(vty, "no test in progress\n");
+	}
+	return CMD_SUCCESS;
+}
+
+DEFPY(clear_labelpool_perf_test, clear_labelpool_perf_test_cmd,
+      "debug bgp lptest clear",
+      DEBUG_STR BGP_STR
+      "label pool test\n"
+      "clear\n")
+{
+
+	if (lpt_inprogress) {
+		lptest_stop();
+	}
+	if (lp_tests) {
+		while (!skiplist_first(lp_tests, NULL, NULL))
+			/* del function of skiplist cleans up tcbs */
+			skiplist_delete_first(lp_tests);
+	}
+	return CMD_SUCCESS;
+}
+
+/*
+ * With the "release" command, we can release labels at intervals through
+ * the ID space. Thus we can to exercise the bitfield-wrapping behavior
+ * of the allocator in a subsequent test.
+ */
+/* clang-format off */
+DEFPY(release_labelpool_perf_test, release_labelpool_perf_test_cmd,
+      "debug bgp lptest release test GENERATION$generation every (1-5)$every_nth",
+      DEBUG_STR
+      BGP_STR
+      "label pool test\n"
+      "release labels\n"
+      "\"test\"\n"
+      "test number\n"
+      "\"every\"\n"
+      "label fraction denominator\n")
+{
+	/* clang-format on */
+
+	unsigned long testnum;
+	char *end;
+	struct lp_test *tcb;
+
+	testnum = strtoul(generation, &end, 0);
+	if (*end) {
+		vty_out(vty, "Invalid test number: \"%s\"\n", generation);
+		return CMD_SUCCESS;
+	}
+	if (lpt_inprogress && (testnum == lpt_generation)) {
+		vty_out(vty,
+			"Error: Test %lu is still in progress (stop first)\n",
+			testnum);
+		return CMD_SUCCESS;
+	}
+
+	if (skiplist_search(lp_tests, (void *)(uintptr_t)testnum,
+			    (void **)&tcb)) {
+
+		/* couldn't find current test in progress */
+		vty_out(vty, "Error: Can't look up test number: \"%lu\"\n",
+			testnum);
+		++lpt_release_tcb_lookup_fails;
+		return CMD_SUCCESS;
+	}
+
+	void *Key, *cKey;
+	void *Value, *cValue;
+	void *cursor;
+	unsigned int iteration;
+	int rc;
+
+	cursor = NULL;
+	iteration = 0;
+	rc = skiplist_next(tcb->labels, &Key, &Value, &cursor);
+
+	while (!rc) {
+		cKey = Key;
+		cValue = Value;
+
+		/* find next item before we delete this one */
+		rc = skiplist_next(tcb->labels, &Key, &Value, &cursor);
+
+		if (!(iteration % every_nth)) {
+			bgp_lp_release(tcb->label_type, cKey,
+				       (mpls_label_t)(uintptr_t)cValue);
+			skiplist_delete(tcb->labels, cKey, NULL);
+			++tcb->counter[LPT_STAT_DEALLOCATED];
+		}
+		++iteration;
+	}
+
+	return CMD_SUCCESS;
+}
+
+static void lptest_delete(void *val)
+{
+	struct lp_test *tcb = (struct lp_test *)val;
+	void *Key;
+	void *Value;
+	void *cursor;
+
+	if (tcb->labels) {
+		cursor = NULL;
+		while (!skiplist_next(tcb->labels, &Key, &Value, &cursor))
+			bgp_lp_release(tcb->label_type, Key,
+				       (mpls_label_t)(uintptr_t)Value);
+		skiplist_free(tcb->labels);
+		tcb->labels = NULL;
+	}
+	if (tcb->timestamps_alloc) {
+		cursor = NULL;
+		skiplist_free(tcb->timestamps_alloc);
+		tcb->timestamps_alloc = NULL;
+	}
+
+	if (tcb->timestamps_dealloc) {
+		cursor = NULL;
+		skiplist_free(tcb->timestamps_dealloc);
+		tcb->timestamps_dealloc = NULL;
+	}
+
+	if (tcb->event_thread)
+		thread_cancel(&tcb->event_thread);
+
+	memset(tcb, 0, sizeof(*tcb));
+
+	XFREE(MTYPE_LABELPOOL_TEST, tcb);
+}
+
+static void lptest_init(void)
+{
+	lp_tests = skiplist_new(0, NULL, lptest_delete);
+}
+
+static void lptest_finish(void)
+{
+	if (lp_tests) {
+		skiplist_free(lp_tests);
+		lp_tests = NULL;
+	}
+}
+
+/*------------------------------------------------------------------------
+ *			Testing code end
+ *------------------------------------------------------------------------*/
+#endif /* BGP_LABELPOOL_ENABLE_TESTS */
+
 void bgp_lp_vty_init(void)
 {
 	install_element(VIEW_NODE, &show_bgp_labelpool_summary_cmd);
@@ -963,4 +1548,12 @@ void bgp_lp_vty_init(void)
 	install_element(VIEW_NODE, &show_bgp_labelpool_inuse_cmd);
 	install_element(VIEW_NODE, &show_bgp_labelpool_requests_cmd);
 	install_element(VIEW_NODE, &show_bgp_labelpool_chunks_cmd);
+
+#if BGP_LABELPOOL_ENABLE_TESTS
+	install_element(ENABLE_NODE, &start_labelpool_perf_test_cmd);
+	install_element(ENABLE_NODE, &show_labelpool_perf_test_cmd);
+	install_element(ENABLE_NODE, &stop_labelpool_perf_test_cmd);
+	install_element(ENABLE_NODE, &release_labelpool_perf_test_cmd);
+	install_element(ENABLE_NODE, &clear_labelpool_perf_test_cmd);
+#endif /* BGP_LABELPOOL_ENABLE_TESTS */
 }
diff --git a/bgpd/bgp_labelpool.h b/bgpd/bgp_labelpool.h
index d6a8eec84..2f3ffe437 100644
--- a/bgpd/bgp_labelpool.h
+++ b/bgpd/bgp_labelpool.h
@@ -41,6 +41,7 @@ struct labelpool {
 	struct work_queue	*callback_q;
 	uint32_t		pending_count;	/* requested from zebra */
 	uint32_t reconnect_count;		/* zebra reconnections */
+	uint32_t next_chunksize;		/* request this many labels */
 };
 
 extern void bgp_lp_init(struct thread_master *master, struct labelpool *pool);
diff --git a/bgpd/subdir.am b/bgpd/subdir.am
index 9e3a09552..b1eeb937e 100644
--- a/bgpd/subdir.am
+++ b/bgpd/subdir.am
@@ -232,6 +232,7 @@ clippy_scan += \
 	bgpd/bgp_bmp.c \
 	bgpd/bgp_debug.c \
 	bgpd/bgp_evpn_vty.c \
+	bgpd/bgp_labelpool.c \
 	bgpd/bgp_route.c \
 	bgpd/bgp_routemap.c \
 	bgpd/bgp_rpki.c \
diff --git a/lib/bitfield.h b/lib/bitfield.h
index a3f361ed9..9af4053da 100644
--- a/lib/bitfield.h
+++ b/lib/bitfield.h
@@ -158,6 +158,75 @@ DECLARE_MTYPE(BITFIELD);
 		(b) += (w * WORD_SIZE);                                        \
 	} while (0)
 
+/*
+ * Find a clear bit in v and return it
+ * Start looking in the word containing bit position start_index.
+ * If necessary, wrap around after bit position max_index.
+ */
+static inline unsigned int
+bf_find_next_clear_bit_wrap(bitfield_t *v, word_t start_index, word_t max_index)
+{
+	int start_bit;
+	unsigned long i, offset, scanbits, wordcount_max, index_max;
+
+	if (start_index > max_index)
+		start_index = 0;
+
+	start_bit = start_index & (WORD_SIZE - 1);
+	wordcount_max = bf_index(max_index) + 1;
+
+	scanbits = WORD_SIZE;
+	for (i = bf_index(start_index); i < v->m; ++i) {
+		if (v->data[i] == WORD_MAX) {
+			/* if the whole word is full move to the next */
+			start_bit = 0;
+			continue;
+		}
+		/* scan one word for clear bits */
+		if ((i == v->m - 1) && (v->m >= wordcount_max))
+			/* max index could be only part of word */
+			scanbits = (max_index % WORD_SIZE) + 1;
+		for (offset = start_bit; offset < scanbits; ++offset) {
+			if (!((v->data[i] >> offset) & 1))
+				return ((i * WORD_SIZE) + offset);
+		}
+		/* move to the next word */
+		start_bit = 0;
+	}
+
+	if (v->m < wordcount_max) {
+		/*
+		 * We can expand bitfield, so no need to wrap.
+		 * Return the index of the first bit of the next word.
+		 * Assumption is that caller will call bf_set_bit which
+		 * will allocate additional space.
+		 */
+		v->m += 1;
+		v->data = (word_t *)realloc(v->data, v->m * sizeof(word_t));
+		v->data[v->m - 1] = 0;
+		return v->m * WORD_SIZE;
+	}
+
+	/*
+	 * start looking for a clear bit at the start of the bitfield and
+	 * stop when we reach start_index
+	 */
+	scanbits = WORD_SIZE;
+	index_max = bf_index(start_index - 1);
+	for (i = 0; i <= index_max; ++i) {
+		if (i == index_max)
+			scanbits = ((start_index - 1) % WORD_SIZE) + 1;
+		for (offset = start_bit; offset < scanbits; ++offset) {
+			if (!((v->data[i] >> offset) & 1))
+				return ((i * WORD_SIZE) + offset);
+		}
+		/* move to the next word */
+		start_bit = 0;
+	}
+
+	return WORD_MAX;
+}
+
 static inline unsigned int bf_find_next_set_bit(bitfield_t v,
 		word_t start_index)
 {
diff --git a/tests/topotests/bgp_lu_topo1/R1/labelpool.summ.json b/tests/topotests/bgp_lu_topo1/R1/labelpool.summ.json
index 29e6c2cbf..e2eee513e 100644
--- a/tests/topotests/bgp_lu_topo1/R1/labelpool.summ.json
+++ b/tests/topotests/bgp_lu_topo1/R1/labelpool.summ.json
@@ -2,7 +2,7 @@
   "Ledger":506,
   "InUse":506,
   "Requests":0,
-  "LabelChunks":11,
+  "LabelChunks":3,
   "Pending":0,
   "Reconnects":0
 }
diff --git a/tests/topotests/bgp_lu_topo2/R1/labelpool.summ.json b/tests/topotests/bgp_lu_topo2/R1/labelpool.summ.json
index 77705e8d3..0dc59b58c 100644
--- a/tests/topotests/bgp_lu_topo2/R1/labelpool.summ.json
+++ b/tests/topotests/bgp_lu_topo2/R1/labelpool.summ.json
@@ -2,7 +2,7 @@
   "Ledger":51,
   "InUse":51,
   "Requests":0,
-  "LabelChunks":2,
+  "LabelChunks":1,
   "Pending":0,
   "Reconnects":0
 }