diff --git a/lib/ns/client.c b/lib/ns/client.c
index f13ff04c76..58d5c48624 100644
--- a/lib/ns/client.c
+++ b/lib/ns/client.c
@@ -334,7 +334,7 @@ client_allocsendbuf(ns_client_t *client, isc_buffer_t *buffer,
 
 	if (TCP_CLIENT(client)) {
 		INSIST(client->tcpbuf == NULL);
-		client->tcpbuf = isc_mem_get(client->manager->mctx,
+		client->tcpbuf = isc_mem_get(client->manager->send_mctx,
 					     NS_CLIENT_TCP_BUFFER_SIZE);
 		client->tcpbuf_size = NS_CLIENT_TCP_BUFFER_SIZE;
 		data = client->tcpbuf;
@@ -371,9 +371,9 @@ client_sendpkg(ns_client_t *client, isc_buffer_t *buffer) {
 
 	if (isc_buffer_base(buffer) == client->tcpbuf) {
 		size_t used = isc_buffer_usedlength(buffer);
-		client->tcpbuf =
-			isc_mem_creget(client->manager->mctx, client->tcpbuf,
-				       client->tcpbuf_size, used, sizeof(char));
+		client->tcpbuf = isc_mem_creget(
+			client->manager->send_mctx, client->tcpbuf,
+			client->tcpbuf_size, used, sizeof(char));
 		client->tcpbuf_size = used;
 		r.base = client->tcpbuf;
 		r.length = used;
@@ -449,7 +449,7 @@ ns_client_sendraw(ns_client_t *client, dns_message_t *message) {
 	return;
 done:
 	if (client->tcpbuf != NULL) {
-		isc_mem_put(client->manager->mctx, client->tcpbuf,
+		isc_mem_put(client->manager->send_mctx, client->tcpbuf,
 			    client->tcpbuf_size);
 	}
 
@@ -733,7 +733,7 @@ renderend:
 
 cleanup:
 	if (client->tcpbuf != NULL) {
-		isc_mem_put(client->manager->mctx, client->tcpbuf,
+		isc_mem_put(client->manager->send_mctx, client->tcpbuf,
 			    client->tcpbuf_size);
 	}
 
@@ -1621,7 +1621,7 @@ ns__client_reset_cb(void *client0) {
 
 	ns_client_endrequest(client);
 	if (client->tcpbuf != NULL) {
-		isc_mem_put(client->manager->mctx, client->tcpbuf,
+		isc_mem_put(client->manager->send_mctx, client->tcpbuf,
 			    client->tcpbuf_size);
 	}
 
@@ -1658,7 +1658,8 @@ ns__client_put_cb(void *client0) {
 
 	client->magic = 0;
 
-	isc_mem_put(manager->mctx, client->sendbuf, NS_CLIENT_SEND_BUFFER_SIZE);
+	isc_mem_put(manager->send_mctx, client->sendbuf,
+		    NS_CLIENT_SEND_BUFFER_SIZE);
 	if (client->opt != NULL) {
 		INSIST(dns_rdataset_isassociated(client->opt));
 		dns_rdataset_disassociate(client->opt);
@@ -2308,7 +2309,7 @@ ns__client_setup(ns_client_t *client, ns_clientmgr_t *mgr, bool new) {
 		dns_message_create(client->manager->mctx,
 				   DNS_MESSAGE_INTENTPARSE, &client->message);
 
-		client->sendbuf = isc_mem_get(client->manager->mctx,
+		client->sendbuf = isc_mem_get(client->manager->send_mctx,
 					      NS_CLIENT_SEND_BUFFER_SIZE);
 		/*
 		 * Set magic earlier than usual because ns_query_init()
@@ -2356,7 +2357,7 @@ ns__client_setup(ns_client_t *client, ns_clientmgr_t *mgr, bool new) {
 
 cleanup:
 	if (client->sendbuf != NULL) {
-		isc_mem_put(client->manager->mctx, client->sendbuf,
+		isc_mem_put(client->manager->send_mctx, client->sendbuf,
 			    NS_CLIENT_SEND_BUFFER_SIZE);
 	}
 
@@ -2392,6 +2393,8 @@ clientmgr_destroy_cb(void *arg) {
 
 	ns_server_detach(&manager->sctx);
 
+	isc_mem_detach(&manager->send_mctx);
+
 	isc_mem_putanddetach(&manager->mctx, manager, sizeof(*manager));
 }
 
@@ -2424,6 +2427,61 @@ ns_clientmgr_create(ns_server_t *sctx, isc_loopmgr_t *loopmgr,
 	isc_refcount_init(&manager->references, 1);
 	ns_server_attach(sctx, &manager->sctx);
 
+	/*
+	 * We create specialised per-worker memory context specifically
+	 * dedicated and tuned for allocating send buffers as it is a very
+	 * common operation. Not doing so may result in excessive memory
+	 * use in certain workloads.
+	 *
+	 * Please see this thread for more details:
+	 *
+	 * https://github.com/jemalloc/jemalloc/issues/2483
+	 *
+	 * In particular, this information from the jemalloc developers is
+	 * of the most interest:
+	 *
+	 * https://github.com/jemalloc/jemalloc/issues/2483#issuecomment-1639019699
+	 * https://github.com/jemalloc/jemalloc/issues/2483#issuecomment-1698173849
+	 *
+	 * In essence, we use the following memory management strategy:
+	 *
+	 * 1. We use a per-worker memory arena for send buffers memory
+	 * allocation to reduce lock contention (In reality, we create a
+	 * per-client manager arena, but we have one client manager per
+	 * worker).
+	 *
+	 * 2. The automatically created arenas settings remain unchanged
+	 * and may be controlled by users (e.g. by setting the
+	 * "MALLOC_CONF" variable).
+	 *
+	 * 3. We attune the arenas to not use dirty pages cache as the
+	 * cache would have a poor reuse rate, and that is known to
+	 * significantly contribute to excessive memory use.
+	 *
+	 * 4. There is no strict need for the dirty cache, as there is a
+	 * per arena bin for each allocation size, so because we initially
+	 * allocate strictly 64K per send buffer (enough for a DNS
+	 * message), allocations would get directed to one bin (an "object
+	 * pool" or a "slab") maintained within an arena. That is, there
+	 * is an object pool already, specifically to optimise for the
+	 * case of frequent allocations of objects of the given size. The
+	 * object pool should suffice our needs, as we will end up
+	 * recycling the objects from there without the need to back it by
+	 * an additional layer of dirty pages cache. The dirty pages cache
+	 * would have worked better in the case when there are more
+	 * allocation bins involved due to a higher reuse rate (the case
+	 * of a more "generic" memory management).
+	 */
+	isc_mem_create_arena(&manager->send_mctx);
+	isc_mem_setname(manager->send_mctx, "sendbufs");
+	(void)isc_mem_arena_set_dirty_decay_ms(manager->send_mctx, 0);
+	/*
+	 * Disable muzzy pages cache too, as versions < 5.2.0 have it
+	 * enabled by default. The muzzy pages cache goes right below the
+	 * dirty pages cache and backs it.
+	 */
+	(void)isc_mem_arena_set_muzzy_decay_ms(manager->send_mctx, 0);
+
 	manager->magic = MANAGER_MAGIC;
 
 	MTRACE("create");
diff --git a/lib/ns/include/ns/client.h b/lib/ns/include/ns/client.h
index 111e107d58..43cd72d1c8 100644
--- a/lib/ns/include/ns/client.h
+++ b/lib/ns/include/ns/client.h
@@ -144,6 +144,7 @@ struct ns_clientmgr {
 	unsigned int magic;
 
 	isc_mem_t     *mctx;
+	isc_mem_t     *send_mctx;
 	ns_server_t   *sctx;
 	isc_refcount_t references;
 	uint32_t       tid;
diff --git a/lib/ns/server.c b/lib/ns/server.c
index bd17a9baaa..d853848a76 100644
--- a/lib/ns/server.c
+++ b/lib/ns/server.c
@@ -55,6 +55,11 @@ ns_server_create(isc_mem_t *mctx, ns_matchview_t matchingview,
 
 	isc_mem_attach(mctx, &sctx->mctx);
 
+	/*
+	 * See here for more details:
+	 * https://github.com/jemalloc/jemalloc/issues/2483
+	 */
+
 	isc_refcount_init(&sctx->references, 1);
 
 	isc_quota_init(&sctx->xfroutquota, 10);