Skip to contentSkip to conversation

net: openvswitch: reorder masks array based on usage

PATCH Applied Eelco Chaudron 3 participants 3 messages Jul 15, 2020 – Jul 17, 2020

Subsystems (3)
lore
Share & tools
lore

Eelco Chaudron

author maintainer
#0 lore raw

PATCH 1/1 net: openvswitch: reorder masks array based on usage

This patch reorders the masks array every 4 seconds based on their usage count. This greatly reduces the masks per packet hit, and hence the overall performance. Especially in the OVS/OVN case for OpenShift.

Here are some results from the OVS/OVN OpenShift test, which use 8 pods, each pod having 512 uperf connections, each connection sends a 64-byte request and gets a 1024-byte response (TCP). All uperf clients are on 1 worker node while all uperf servers are on the other worker node.

Kernel without this patch : 7.71 Gbps
Kernel with this patch applied: 14.52 Gbps

We also run some tests to verify the rebalance activity does not lower the flow insertion rate, which does not.

---
 net/openvswitch/datapath.c   |   22 +++++
 net/openvswitch/datapath.h   |    8 +-
 net/openvswitch/flow_table.c |  173 +++++++++++++++++++++++++++++++++++++++++-
 net/openvswitch/flow_table.h |   11 +++
 4 files changed, 207 insertions(+), 7 deletions(-)
net/openvswitch/datapath.c +22 -0
index 94b024534987..95805f0e27bd 100644
@@ -130,6 +130,8 @@ static int queue_userspace_packet(struct datapath *dp, struct sk_buff *,
 				  const struct dp_upcall_info *,
 				  uint32_t cutlen);
 
+static void ovs_dp_masks_rebalance(struct work_struct *work);
+
 /* Must be called with rcu_read_lock or ovs_mutex. */
 const char *ovs_dp_name(const struct datapath *dp)
 {
@@ -1653,6 +1655,7 @@ static int ovs_dp_cmd_new(struct sk_buff *skb, struct genl_info *info)
 		goto err_destroy_reply;
 
 	ovs_dp_set_net(dp, sock_net(skb->sk));
+	INIT_DELAYED_WORK(&dp->masks_rebalance, ovs_dp_masks_rebalance);
 
 	/* Allocate table. */
 	err = ovs_flow_tbl_init(&dp->table);
@@ -1712,6 +1715,9 @@ static int ovs_dp_cmd_new(struct sk_buff *skb, struct genl_info *info)
 	ovs_net = net_generic(ovs_dp_get_net(dp), ovs_net_id);
 	list_add_tail_rcu(&dp->list_node, &ovs_net->dps);
 
+	schedule_delayed_work(&dp->masks_rebalance,
+			      msecs_to_jiffies(DP_MASKS_REBALANCE_INTERVAL));
+
 	ovs_unlock();
 
 	ovs_notify(&dp_datapath_genl_family, reply, info);
@@ -1756,6 +1762,9 @@ static void __dp_destroy(struct datapath *dp)
 
 	/* RCU destroy the flow table */
 	call_rcu(&dp->rcu, destroy_dp_rcu);
+
+	/* Cancel remaining work. */
+	cancel_delayed_work_sync(&dp->masks_rebalance);
 }
 
 static int ovs_dp_cmd_del(struct sk_buff *skb, struct genl_info *info)
@@ -2338,6 +2347,19 @@ static int ovs_vport_cmd_dump(struct sk_buff *skb, struct netlink_callback *cb)
 	return skb->len;
 }
 
+static void ovs_dp_masks_rebalance(struct work_struct *work)
+{
+	struct datapath *dp = container_of(work, struct datapath,
+					   masks_rebalance.work);
+
+	ovs_lock();
+	ovs_flow_masks_rebalance(&dp->table);
+	ovs_unlock();
+
+	schedule_delayed_work(&dp->masks_rebalance,
+			      msecs_to_jiffies(DP_MASKS_REBALANCE_INTERVAL));
+}
+
 static const struct nla_policy vport_policy[OVS_VPORT_ATTR_MAX + 1] = {
 	[OVS_VPORT_ATTR_NAME] = { .type = NLA_NUL_STRING, .len = IFNAMSIZ - 1 },
 	[OVS_VPORT_ATTR_STATS] = { .len = sizeof(struct ovs_vport_stats) },
net/openvswitch/datapath.h +6 -2
index 2016dd107939..697a2354194b 100644
@@ -20,8 +20,9 @@
 #include "meter.h"
 #include "vport-internal_dev.h"
 
-#define DP_MAX_PORTS           USHRT_MAX
-#define DP_VPORT_HASH_BUCKETS  1024
+#define DP_MAX_PORTS                USHRT_MAX
+#define DP_VPORT_HASH_BUCKETS       1024
+#define DP_MASKS_REBALANCE_INTERVAL 4000
 
 /**
  * struct dp_stats_percpu - per-cpu packet processing statistics for a given
@@ -83,6 +84,9 @@ struct datapath {
 
 	/* Switch meters. */
 	struct dp_meter_table meter_tbl;
+
+	/* re-balance flow masks timer */
+	struct delayed_work masks_rebalance;
 };
 
 /**
net/openvswitch/flow_table.c +168 -5

252 lines not on this page · raw message

net/openvswitch/flow_table.h +11 -0
index 8a5cea6ae111..1f664b050e3b 100644
@@ -27,9 +27,17 @@ struct mask_cache_entry {
 	u32 mask_index;
 };
 
+struct mask_count {
+	int index;
+	u64 counter;
+};
+
 struct mask_array {
 	struct rcu_head rcu;
 	int count, max;
+	u64 __percpu *masks_usage_cntr;
+	u64 *masks_usage_zero_cntr;
+	struct u64_stats_sync syncp;
 	struct sw_flow_mask __rcu *masks[];
 };
 
@@ -86,4 +94,7 @@ bool ovs_flow_cmp(const struct sw_flow *, const struct sw_flow_match *);
 
 void ovs_flow_mask_key(struct sw_flow_key *dst, const struct sw_flow_key *src,
 		       bool full, const struct sw_flow_mask *mask);
+
+void ovs_flow_masks_rebalance(struct flow_table *table);
+
 #endif /* flow_table.h */
maintainer Reviewed-by on v1 1/1
#1 lore raw
Quoted 11 lines · Eelco Chaudron

This patch reorders the masks array every 4 seconds based on their usage count. This greatly reduces the masks per packet hit, and hence the overall performance. Especially in the OVS/OVN case for OpenShift.

Here are some results from the OVS/OVN OpenShift test, which use 8 pods, each pod having 512 uperf connections, each connection sends a 64-byte request and gets a 1024-byte response (TCP). All uperf clients are on 1 worker node while all uperf servers are on the other worker node.

Quoted 8 lines

Kernel without this patch : 7.71 Gbps
Kernel with this patch applied: 14.52 Gbps

We also run some tests to verify the rebalance activity does not lower the flow insertion rate, which does not.

/P

maintainer Applied
#2 lore raw
In reply to Eelco Chaudron

From: Eelco Chaudron <[email protected]>
Date: Wed, 15 Jul 2020 14:09:28 +0200

Quoted 11 lines

This patch reorders the masks array every 4 seconds based on their usage count. This greatly reduces the masks per packet hit, and hence the overall performance. Especially in the OVS/OVN case for OpenShift.

Here are some results from the OVS/OVN OpenShift test, which use 8 pods, each pod having 512 uperf connections, each connection sends a 64-byte request and gets a 1024-byte response (TCP). All uperf clients are on 1 worker node while all uperf servers are on the other worker node.

Quoted 8 lines

Kernel without this patch : 7.71 Gbps
Kernel with this patch applied: 14.52 Gbps

We also run some tests to verify the rebalance activity does not lower the flow insertion rate, which does not.

Applied, thanks.