* [PATCH v2 1/2] openvswitch: fix flow stats accounting when node 0 is not possible
@ 2016-09-15 22:11 Thadeu Lima de Souza Cascardo
2016-09-15 22:11 ` [PATCH v2 2/2] openvswitch: use percpu flow stats Thadeu Lima de Souza Cascardo
[not found] ` <1473977513-7617-1-git-send-email-cascardo-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
0 siblings, 2 replies; 8+ messages in thread
From: Thadeu Lima de Souza Cascardo @ 2016-09-15 22:11 UTC (permalink / raw
To: netdev-u79uwXL29TY76Z2rM5mHXA
Cc: dev-yBygre7rU0TnMu66kgdUjQ, David Miller, Eric Dumazet
On a system with only node 1 as possible, all statistics is going to be
accounted on node 0 as it will have a single writer.
However, when getting and clearing the statistics, node 0 is not going
to be considered, as it's not a possible node.
Tested that statistics are not zero on a system with only node 1
possible. Also compile-tested with CONFIG_NUMA off.
Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@redhat.com>
---
net/openvswitch/flow.c | 6 ++++--
net/openvswitch/flow_table.c | 5 +++--
2 files changed, 7 insertions(+), 4 deletions(-)
diff --git a/net/openvswitch/flow.c b/net/openvswitch/flow.c
index 1240ae3..5b80612 100644
--- a/net/openvswitch/flow.c
+++ b/net/openvswitch/flow.c
@@ -142,7 +142,8 @@ void ovs_flow_stats_get(const struct sw_flow *flow,
*tcp_flags = 0;
memset(ovs_stats, 0, sizeof(*ovs_stats));
- for_each_node(node) {
+ /* We open code this to make sure node 0 is always considered */
+ for (node = 0; node < MAX_NUMNODES; node = next_node(node, node_possible_map)) {
struct flow_stats *stats = rcu_dereference_ovsl(flow->stats[node]);
if (stats) {
@@ -165,7 +166,8 @@ void ovs_flow_stats_clear(struct sw_flow *flow)
{
int node;
- for_each_node(node) {
+ /* We open code this to make sure node 0 is always considered */
+ for (node = 0; node < MAX_NUMNODES; node = next_node(node, node_possible_map)) {
struct flow_stats *stats = ovsl_dereference(flow->stats[node]);
if (stats) {
diff --git a/net/openvswitch/flow_table.c b/net/openvswitch/flow_table.c
index d073fff..957a3c3 100644
--- a/net/openvswitch/flow_table.c
+++ b/net/openvswitch/flow_table.c
@@ -148,8 +148,9 @@ static void flow_free(struct sw_flow *flow)
kfree(flow->id.unmasked_key);
if (flow->sf_acts)
ovs_nla_free_flow_actions((struct sw_flow_actions __force *)flow->sf_acts);
- for_each_node(node)
- if (flow->stats[node])
+ /* We open code this to make sure node 0 is always considered */
+ for (node = 0; node < MAX_NUMNODES; node = next_node(node, node_possible_map))
+ if (node != 0 && flow->stats[node])
kmem_cache_free(flow_stats_cache,
(struct flow_stats __force *)flow->stats[node]);
kmem_cache_free(flow_cache, flow);
--
2.7.4
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
^ permalink raw reply related [flat|nested] 8+ messages in thread
* [PATCH v2 2/2] openvswitch: use percpu flow stats
2016-09-15 22:11 [PATCH v2 1/2] openvswitch: fix flow stats accounting when node 0 is not possible Thadeu Lima de Souza Cascardo
@ 2016-09-15 22:11 ` Thadeu Lima de Souza Cascardo
2016-09-15 23:09 ` Eric Dumazet
[not found] ` <1473977513-7617-2-git-send-email-cascardo-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
[not found] ` <1473977513-7617-1-git-send-email-cascardo-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
1 sibling, 2 replies; 8+ messages in thread
From: Thadeu Lima de Souza Cascardo @ 2016-09-15 22:11 UTC (permalink / raw
To: netdev; +Cc: dev, pshelar, David Miller, Eric Dumazet
Instead of using flow stats per NUMA node, use it per CPU. When using
megaflows, the stats lock can be a bottleneck in scalability.
On a E5-2690 12-core system, usual throughput went from ~4Mpps to
~15Mpps when forwarding between two 40GbE ports with a single flow
configured on the datapath.
This has been tested on a system with possible CPUs 0-7,16-23. After
module removal, there were no corruption on the slab cache.
Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@redhat.com>
Cc: pravin shelar <pshelar@ovn.org>
---
v2:
* use smp_processor_id as ovs_flow_stats_update is always called from BH
context
* use kmem_cache_zalloc to allocate flow
---
net/openvswitch/flow.c | 42 ++++++++++++++++++++++--------------------
net/openvswitch/flow.h | 4 ++--
net/openvswitch/flow_table.c | 26 +++++++++-----------------
3 files changed, 33 insertions(+), 39 deletions(-)
diff --git a/net/openvswitch/flow.c b/net/openvswitch/flow.c
index 5b80612..0fa45439 100644
--- a/net/openvswitch/flow.c
+++ b/net/openvswitch/flow.c
@@ -29,6 +29,7 @@
#include <linux/module.h>
#include <linux/in.h>
#include <linux/rcupdate.h>
+#include <linux/cpumask.h>
#include <linux/if_arp.h>
#include <linux/ip.h>
#include <linux/ipv6.h>
@@ -72,32 +73,33 @@ void ovs_flow_stats_update(struct sw_flow *flow, __be16 tcp_flags,
{
struct flow_stats *stats;
int node = numa_node_id();
+ int cpu = smp_processor_id();
int len = skb->len + (skb_vlan_tag_present(skb) ? VLAN_HLEN : 0);
- stats = rcu_dereference(flow->stats[node]);
+ stats = rcu_dereference(flow->stats[cpu]);
- /* Check if already have node-specific stats. */
+ /* Check if already have CPU-specific stats. */
if (likely(stats)) {
spin_lock(&stats->lock);
/* Mark if we write on the pre-allocated stats. */
- if (node == 0 && unlikely(flow->stats_last_writer != node))
- flow->stats_last_writer = node;
+ if (cpu == 0 && unlikely(flow->stats_last_writer != cpu))
+ flow->stats_last_writer = cpu;
} else {
stats = rcu_dereference(flow->stats[0]); /* Pre-allocated. */
spin_lock(&stats->lock);
- /* If the current NUMA-node is the only writer on the
+ /* If the current CPU is the only writer on the
* pre-allocated stats keep using them.
*/
- if (unlikely(flow->stats_last_writer != node)) {
+ if (unlikely(flow->stats_last_writer != cpu)) {
/* A previous locker may have already allocated the
- * stats, so we need to check again. If node-specific
+ * stats, so we need to check again. If CPU-specific
* stats were already allocated, we update the pre-
* allocated stats as we have already locked them.
*/
- if (likely(flow->stats_last_writer != NUMA_NO_NODE)
- && likely(!rcu_access_pointer(flow->stats[node]))) {
- /* Try to allocate node-specific stats. */
+ if (likely(flow->stats_last_writer != -1) &&
+ likely(!rcu_access_pointer(flow->stats[cpu]))) {
+ /* Try to allocate CPU-specific stats. */
struct flow_stats *new_stats;
new_stats =
@@ -114,12 +116,12 @@ void ovs_flow_stats_update(struct sw_flow *flow, __be16 tcp_flags,
new_stats->tcp_flags = tcp_flags;
spin_lock_init(&new_stats->lock);
- rcu_assign_pointer(flow->stats[node],
+ rcu_assign_pointer(flow->stats[cpu],
new_stats);
goto unlock;
}
}
- flow->stats_last_writer = node;
+ flow->stats_last_writer = cpu;
}
}
@@ -136,15 +138,15 @@ void ovs_flow_stats_get(const struct sw_flow *flow,
struct ovs_flow_stats *ovs_stats,
unsigned long *used, __be16 *tcp_flags)
{
- int node;
+ int cpu;
*used = 0;
*tcp_flags = 0;
memset(ovs_stats, 0, sizeof(*ovs_stats));
- /* We open code this to make sure node 0 is always considered */
- for (node = 0; node < MAX_NUMNODES; node = next_node(node, node_possible_map)) {
- struct flow_stats *stats = rcu_dereference_ovsl(flow->stats[node]);
+ /* We open code this to make sure cpu 0 is always considered */
+ for (cpu = 0; cpu < nr_cpu_ids; cpu = cpumask_next(cpu, cpu_possible_mask)) {
+ struct flow_stats *stats = rcu_dereference_ovsl(flow->stats[cpu]);
if (stats) {
/* Local CPU may write on non-local stats, so we must
@@ -164,11 +166,11 @@ void ovs_flow_stats_get(const struct sw_flow *flow,
/* Called with ovs_mutex. */
void ovs_flow_stats_clear(struct sw_flow *flow)
{
- int node;
+ int cpu;
- /* We open code this to make sure node 0 is always considered */
- for (node = 0; node < MAX_NUMNODES; node = next_node(node, node_possible_map)) {
- struct flow_stats *stats = ovsl_dereference(flow->stats[node]);
+ /* We open code this to make sure cpu 0 is always considered */
+ for (cpu = 0; cpu < nr_cpu_ids; cpu = cpumask_next(cpu, cpu_possible_mask)) {
+ struct flow_stats *stats = ovsl_dereference(flow->stats[cpu]);
if (stats) {
spin_lock_bh(&stats->lock);
diff --git a/net/openvswitch/flow.h b/net/openvswitch/flow.h
index 156a302..ae783f5 100644
--- a/net/openvswitch/flow.h
+++ b/net/openvswitch/flow.h
@@ -178,14 +178,14 @@ struct sw_flow {
struct hlist_node node[2];
u32 hash;
} flow_table, ufid_table;
- int stats_last_writer; /* NUMA-node id of the last writer on
+ int stats_last_writer; /* CPU id of the last writer on
* 'stats[0]'.
*/
struct sw_flow_key key;
struct sw_flow_id id;
struct sw_flow_mask *mask;
struct sw_flow_actions __rcu *sf_acts;
- struct flow_stats __rcu *stats[]; /* One for each NUMA node. First one
+ struct flow_stats __rcu *stats[]; /* One for each CPU. First one
* is allocated at flow creation time,
* the rest are allocated on demand
* while holding the 'stats[0].lock'.
diff --git a/net/openvswitch/flow_table.c b/net/openvswitch/flow_table.c
index 957a3c3..ea7a807 100644
--- a/net/openvswitch/flow_table.c
+++ b/net/openvswitch/flow_table.c
@@ -32,6 +32,7 @@
#include <linux/module.h>
#include <linux/in.h>
#include <linux/rcupdate.h>
+#include <linux/cpumask.h>
#include <linux/if_arp.h>
#include <linux/ip.h>
#include <linux/ipv6.h>
@@ -79,17 +80,12 @@ struct sw_flow *ovs_flow_alloc(void)
{
struct sw_flow *flow;
struct flow_stats *stats;
- int node;
- flow = kmem_cache_alloc(flow_cache, GFP_KERNEL);
+ flow = kmem_cache_zalloc(flow_cache, GFP_KERNEL);
if (!flow)
return ERR_PTR(-ENOMEM);
- flow->sf_acts = NULL;
- flow->mask = NULL;
- flow->id.unmasked_key = NULL;
- flow->id.ufid_len = 0;
- flow->stats_last_writer = NUMA_NO_NODE;
+ flow->stats_last_writer = -1;
/* Initialize the default stat node. */
stats = kmem_cache_alloc_node(flow_stats_cache,
@@ -102,10 +98,6 @@ struct sw_flow *ovs_flow_alloc(void)
RCU_INIT_POINTER(flow->stats[0], stats);
- for_each_node(node)
- if (node != 0)
- RCU_INIT_POINTER(flow->stats[node], NULL);
-
return flow;
err:
kmem_cache_free(flow_cache, flow);
@@ -142,17 +134,17 @@ static struct flex_array *alloc_buckets(unsigned int n_buckets)
static void flow_free(struct sw_flow *flow)
{
- int node;
+ int cpu;
if (ovs_identifier_is_key(&flow->id))
kfree(flow->id.unmasked_key);
if (flow->sf_acts)
ovs_nla_free_flow_actions((struct sw_flow_actions __force *)flow->sf_acts);
- /* We open code this to make sure node 0 is always considered */
- for (node = 0; node < MAX_NUMNODES; node = next_node(node, node_possible_map))
- if (node != 0 && flow->stats[node])
+ /* We open code this to make sure cpu 0 is always considered */
+ for (cpu = 0; cpu < nr_cpu_ids; cpu = cpumask_next(cpu, cpu_possible_mask))
+ if (flow->stats[cpu])
kmem_cache_free(flow_stats_cache,
- (struct flow_stats __force *)flow->stats[node]);
+ (struct flow_stats __force *)flow->stats[cpu]);
kmem_cache_free(flow_cache, flow);
}
@@ -757,7 +749,7 @@ int ovs_flow_init(void)
BUILD_BUG_ON(sizeof(struct sw_flow_key) % sizeof(long));
flow_cache = kmem_cache_create("sw_flow", sizeof(struct sw_flow)
- + (nr_node_ids
+ + (nr_cpu_ids
* sizeof(struct flow_stats *)),
0, 0, NULL);
if (flow_cache == NULL)
--
2.7.4
^ permalink raw reply related [flat|nested] 8+ messages in thread
* Re: [PATCH v2 2/2] openvswitch: use percpu flow stats
2016-09-15 22:11 ` [PATCH v2 2/2] openvswitch: use percpu flow stats Thadeu Lima de Souza Cascardo
@ 2016-09-15 23:09 ` Eric Dumazet
[not found] ` <1473980966.22679.58.camel-XN9IlZ5yJG9HTL0Zs8A6p+yfmBU6pStAUsxypvmhUTTZJqsBc5GL+g@public.gmane.org>
[not found] ` <1473977513-7617-2-git-send-email-cascardo-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
1 sibling, 1 reply; 8+ messages in thread
From: Eric Dumazet @ 2016-09-15 23:09 UTC (permalink / raw
To: Thadeu Lima de Souza Cascardo; +Cc: netdev, dev, pshelar, David Miller
On Thu, 2016-09-15 at 19:11 -0300, Thadeu Lima de Souza Cascardo wrote:
> Instead of using flow stats per NUMA node, use it per CPU. When using
> megaflows, the stats lock can be a bottleneck in scalability.
>
> On a E5-2690 12-core system, usual throughput went from ~4Mpps to
> ~15Mpps when forwarding between two 40GbE ports with a single flow
> configured on the datapath.
>
> This has been tested on a system with possible CPUs 0-7,16-23. After
> module removal, there were no corruption on the slab cache.
>
> Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@redhat.com>
> Cc: pravin shelar <pshelar@ovn.org>
> ---
> + /* We open code this to make sure cpu 0 is always considered */
> + for (cpu = 0; cpu < nr_cpu_ids; cpu = cpumask_next(cpu, cpu_possible_mask))
> + if (flow->stats[cpu])
> kmem_cache_free(flow_stats_cache,
> - (struct flow_stats __force *)flow->stats[node]);
> + (struct flow_stats __force *)flow->stats[cpu]);
> kmem_cache_free(flow_cache, flow);
> }
>
> @@ -757,7 +749,7 @@ int ovs_flow_init(void)
> BUILD_BUG_ON(sizeof(struct sw_flow_key) % sizeof(long));
>
> flow_cache = kmem_cache_create("sw_flow", sizeof(struct sw_flow)
> - + (nr_node_ids
> + + (nr_cpu_ids
> * sizeof(struct flow_stats *)),
> 0, 0, NULL);
> if (flow_cache == NULL)
Well, if you switch to percpu stats, better use normal
alloc_percpu(struct flow_stats)
The code was dealing with per node allocation so could not use existing
helper.
No need to keep this forever.
^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [PATCH v2 2/2] openvswitch: use percpu flow stats
[not found] ` <1473980966.22679.58.camel-XN9IlZ5yJG9HTL0Zs8A6p+yfmBU6pStAUsxypvmhUTTZJqsBc5GL+g@public.gmane.org>
@ 2016-09-15 23:36 ` Thadeu Lima de Souza Cascardo
0 siblings, 0 replies; 8+ messages in thread
From: Thadeu Lima de Souza Cascardo @ 2016-09-15 23:36 UTC (permalink / raw
To: Eric Dumazet
Cc: dev-yBygre7rU0TnMu66kgdUjQ, netdev-u79uwXL29TY76Z2rM5mHXA,
David Miller
On Thu, Sep 15, 2016 at 04:09:26PM -0700, Eric Dumazet wrote:
> On Thu, 2016-09-15 at 19:11 -0300, Thadeu Lima de Souza Cascardo wrote:
> > Instead of using flow stats per NUMA node, use it per CPU. When using
> > megaflows, the stats lock can be a bottleneck in scalability.
> >
> > On a E5-2690 12-core system, usual throughput went from ~4Mpps to
> > ~15Mpps when forwarding between two 40GbE ports with a single flow
> > configured on the datapath.
> >
> > This has been tested on a system with possible CPUs 0-7,16-23. After
> > module removal, there were no corruption on the slab cache.
> >
> > Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@redhat.com>
> > Cc: pravin shelar <pshelar@ovn.org>
> > ---
>
> > + /* We open code this to make sure cpu 0 is always considered */
> > + for (cpu = 0; cpu < nr_cpu_ids; cpu = cpumask_next(cpu, cpu_possible_mask))
> > + if (flow->stats[cpu])
> > kmem_cache_free(flow_stats_cache,
> > - (struct flow_stats __force *)flow->stats[node]);
> > + (struct flow_stats __force *)flow->stats[cpu]);
> > kmem_cache_free(flow_cache, flow);
> > }
> >
> > @@ -757,7 +749,7 @@ int ovs_flow_init(void)
> > BUILD_BUG_ON(sizeof(struct sw_flow_key) % sizeof(long));
> >
> > flow_cache = kmem_cache_create("sw_flow", sizeof(struct sw_flow)
> > - + (nr_node_ids
> > + + (nr_cpu_ids
> > * sizeof(struct flow_stats *)),
> > 0, 0, NULL);
> > if (flow_cache == NULL)
>
> Well, if you switch to percpu stats, better use normal
> alloc_percpu(struct flow_stats)
>
> The code was dealing with per node allocation so could not use existing
> helper.
>
> No need to keep this forever.
The problem is that the alloc_percpu uses a global spinlock and that affects
some workloads on OVS that creates lots of flows, as described in commit
9ac56358dec1a5aa7f4275a42971f55fad1f7f35 ("datapath: Per NUMA node flow
stats.").
This problem would not happen on this version as the flow allocation does not
suffer from the same scalability problem as when using alloc_percpu.
Cascardo.
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [PATCH v2 1/2] openvswitch: fix flow stats accounting when node 0 is not possible
[not found] ` <1473977513-7617-1-git-send-email-cascardo-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
@ 2016-09-18 2:06 ` pravin shelar
2016-09-19 2:14 ` David Miller
1 sibling, 0 replies; 8+ messages in thread
From: pravin shelar @ 2016-09-18 2:06 UTC (permalink / raw
To: Thadeu Lima de Souza Cascardo
Cc: ovs dev, Linux Kernel Network Developers, David Miller,
Eric Dumazet
On Thu, Sep 15, 2016 at 3:11 PM, Thadeu Lima de Souza Cascardo
<cascardo@redhat.com> wrote:
> On a system with only node 1 as possible, all statistics is going to be
> accounted on node 0 as it will have a single writer.
>
> However, when getting and clearing the statistics, node 0 is not going
> to be considered, as it's not a possible node.
>
> Tested that statistics are not zero on a system with only node 1
> possible. Also compile-tested with CONFIG_NUMA off.
>
> Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@redhat.com>
Acked-by: Pravin B Shelar <pshelar@ovn.org>
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [PATCH v2 2/2] openvswitch: use percpu flow stats
[not found] ` <1473977513-7617-2-git-send-email-cascardo-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
@ 2016-09-18 2:07 ` pravin shelar
2016-09-19 2:14 ` David Miller
1 sibling, 0 replies; 8+ messages in thread
From: pravin shelar @ 2016-09-18 2:07 UTC (permalink / raw
To: Thadeu Lima de Souza Cascardo
Cc: ovs dev, Linux Kernel Network Developers, David Miller,
Eric Dumazet
On Thu, Sep 15, 2016 at 3:11 PM, Thadeu Lima de Souza Cascardo
<cascardo@redhat.com> wrote:
> Instead of using flow stats per NUMA node, use it per CPU. When using
> megaflows, the stats lock can be a bottleneck in scalability.
>
> On a E5-2690 12-core system, usual throughput went from ~4Mpps to
> ~15Mpps when forwarding between two 40GbE ports with a single flow
> configured on the datapath.
>
> This has been tested on a system with possible CPUs 0-7,16-23. After
> module removal, there were no corruption on the slab cache.
>
> Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@redhat.com>
> Cc: pravin shelar <pshelar@ovn.org>
Looks good.
Acked-by: Pravin B Shelar <pshelar@ovn.org>
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [PATCH v2 1/2] openvswitch: fix flow stats accounting when node 0 is not possible
[not found] ` <1473977513-7617-1-git-send-email-cascardo-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
2016-09-18 2:06 ` [PATCH v2 1/2] openvswitch: fix flow stats accounting when node 0 is not possible pravin shelar
@ 2016-09-19 2:14 ` David Miller
1 sibling, 0 replies; 8+ messages in thread
From: David Miller @ 2016-09-19 2:14 UTC (permalink / raw
To: cascardo-H+wXaHxf7aLQT0dZR+AlfA
Cc: dev-yBygre7rU0TnMu66kgdUjQ, netdev-u79uwXL29TY76Z2rM5mHXA,
eric.dumazet-Re5JQEeQqe8AvxtiuMwx3w
From: Thadeu Lima de Souza Cascardo <cascardo@redhat.com>
Date: Thu, 15 Sep 2016 19:11:52 -0300
> On a system with only node 1 as possible, all statistics is going to be
> accounted on node 0 as it will have a single writer.
>
> However, when getting and clearing the statistics, node 0 is not going
> to be considered, as it's not a possible node.
>
> Tested that statistics are not zero on a system with only node 1
> possible. Also compile-tested with CONFIG_NUMA off.
>
> Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@redhat.com>
Applied to net-next.
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [PATCH v2 2/2] openvswitch: use percpu flow stats
[not found] ` <1473977513-7617-2-git-send-email-cascardo-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
2016-09-18 2:07 ` pravin shelar
@ 2016-09-19 2:14 ` David Miller
1 sibling, 0 replies; 8+ messages in thread
From: David Miller @ 2016-09-19 2:14 UTC (permalink / raw
To: cascardo-H+wXaHxf7aLQT0dZR+AlfA
Cc: dev-yBygre7rU0TnMu66kgdUjQ, netdev-u79uwXL29TY76Z2rM5mHXA,
eric.dumazet-Re5JQEeQqe8AvxtiuMwx3w
From: Thadeu Lima de Souza Cascardo <cascardo@redhat.com>
Date: Thu, 15 Sep 2016 19:11:53 -0300
> Instead of using flow stats per NUMA node, use it per CPU. When using
> megaflows, the stats lock can be a bottleneck in scalability.
>
> On a E5-2690 12-core system, usual throughput went from ~4Mpps to
> ~15Mpps when forwarding between two 40GbE ports with a single flow
> configured on the datapath.
>
> This has been tested on a system with possible CPUs 0-7,16-23. After
> module removal, there were no corruption on the slab cache.
>
> Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@redhat.com>
Also applied to net-next, thanks.
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
^ permalink raw reply [flat|nested] 8+ messages in thread
end of thread, other threads:[~2016-09-19 2:14 UTC | newest]
Thread overview: 8+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2016-09-15 22:11 [PATCH v2 1/2] openvswitch: fix flow stats accounting when node 0 is not possible Thadeu Lima de Souza Cascardo
2016-09-15 22:11 ` [PATCH v2 2/2] openvswitch: use percpu flow stats Thadeu Lima de Souza Cascardo
2016-09-15 23:09 ` Eric Dumazet
[not found] ` <1473980966.22679.58.camel-XN9IlZ5yJG9HTL0Zs8A6p+yfmBU6pStAUsxypvmhUTTZJqsBc5GL+g@public.gmane.org>
2016-09-15 23:36 ` Thadeu Lima de Souza Cascardo
[not found] ` <1473977513-7617-2-git-send-email-cascardo-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
2016-09-18 2:07 ` pravin shelar
2016-09-19 2:14 ` David Miller
[not found] ` <1473977513-7617-1-git-send-email-cascardo-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
2016-09-18 2:06 ` [PATCH v2 1/2] openvswitch: fix flow stats accounting when node 0 is not possible pravin shelar
2016-09-19 2:14 ` David Miller
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.