On Thu, Jan 17, 2013 at 09:42:29AM +0900, Simon Horman wrote:
> On Wed, Jan 16, 2013 at 01:59:21PM -0800, Ben Pfaff wrote:
> > On Tue, Jan 15, 2013 at 01:20:57PM +0900, Simon Horman wrote:
> > > Optimise OpenFlow flow expiry by placing expirable flows on a list.
> > > This optimises scanning of flows for expiry in two ways:
> > >
> > > * Empirically list traversal appears faster than the code it replaces.
> > >
> > > With 1,000,000 flows present an otherwise idle system I observed CPU
> > > utilisation of around 20% with the existing code but around 10% with
> > > this new code.
> > >
> > > * Flows that will never expire are not traversed.
> > >
> > > This addresses a case seen in the field.
> >
> > This version looks better. I still have a few comments, but before
> > that, may I ask a little bit about the situation in which the
> > performance improvement was observed? In this situation, about how
> > many of the 1,000,000 flows were actually expirable, that is, had
> > either a hard timeout or an idle timeout? That is, is the performance
> > improvement due more to the first or the second bullet you list above?
> > If none of the flows were expirable, then I guess it was the second;
> > if all of them were, then I guess it was the first; and otherwise it
> > is something in between.
> >
> > Basically I'm wondering if we should do something to make flow table
> > traversal faster, independent of expiration.
>
> Hi Ben,
>
> the primary aim of this patch was to address a performance issue that
> was noticed when inserting 100,000 flows none of which were expirable.
> I have been told this is representative of an expected use-case.
>
> During my testing I used 1,000,000 flows instead of 100,000 in order to
> make the CPU utilisation more pronounced and easier to observe.
>
> In the course of my testing I tested 1,000,000 flows none of which were
> expirable and in that case CPU utilisation was dramatically reduced to
> approximately 0. This seems to be a good outcome for the use-case
> originally reported.
>
> In the course of testing I also tested 1,000,000 flows all of which
> were expirable. This was primarily to see if there were any regressions.
> In the course of this test I noticed that there seemed to be some
> reduction in CPU utilisation. However this was just a side effect and
> not an aim of my work. I should have placed it as the second bullet
> in my commit message and noted that it was a side effect.
OK, I made a few stylistic and comment changes to the patch, and I've
updated the commit message to reflect what you said above. I'm happy
with it now, but I don't want to apply it without your approval since I
changed the meaning of your commit message. Please review?
--8<--------------------------cut here-------------------------->8--
>From 5eee5262555b30c03676975c31543e19e893b1b7 Mon Sep 17 00:00:00 2001
From: Simon Horman <ho...@verge.net.au>
Date: Tue, 15 Jan 2013 13:20:57 +0900
Subject: [PATCH] ofproto: Optimise OpenFlow flow expiry
Optimise OpenFlow flow expiry by placing expirable flows on a list.
This optimises scanning of flows for expiry in two ways:
* Flows that will never expire are not traversed.
This addresses a case seen in the field. With 1,000,000 flows that
are not expirable, this dramatically reduces CPU utilization to
approximately zero.
* Empirically list traversal appears faster than the code it replaces.
With 1,000,000 expirable flows present an otherwise idle system I
observed CPU utilisation of around 20% with the existing code but
around 10% with this new code.
Signed-off-by: Simon Horman <ho...@verge.net.au>
Signed-off-by: Ben Pfaff <b...@nicira.com>
---
ofproto/ofproto-dpif.c | 13 ++++---------
ofproto/ofproto-provider.h | 8 ++++++++
ofproto/ofproto.c | 10 ++++++++++
3 files changed, 22 insertions(+), 9 deletions(-)
diff --git a/ofproto/ofproto-dpif.c b/ofproto/ofproto-dpif.c
index 2c216fe..69da618 100644
--- a/ofproto/ofproto-dpif.c
+++ b/ofproto/ofproto-dpif.c
@@ -3726,8 +3726,7 @@ expire(struct dpif_backer *backer)
update_stats(backer);
HMAP_FOR_EACH (ofproto, all_ofproto_dpifs_node, &all_ofproto_dpifs) {
- struct rule_dpif *rule, *next_rule;
- struct oftable *table;
+ struct rule *rule, *next_rule;
int dp_max_idle;
if (ofproto->backer != backer) {
@@ -3742,13 +3741,9 @@ expire(struct dpif_backer *backer)
/* Expire OpenFlow flows whose idle_timeout or hard_timeout
* has passed. */
- OFPROTO_FOR_EACH_TABLE (table, &ofproto->up) {
- struct cls_cursor cursor;
-
- cls_cursor_init(&cursor, &table->cls, NULL);
- CLS_CURSOR_FOR_EACH_SAFE (rule, next_rule, up.cr, &cursor) {
- rule_expire(rule);
- }
+ LIST_FOR_EACH_SAFE (rule, next_rule, expirable,
+ &ofproto->up.expirable) {
+ rule_expire(rule_dpif_cast(rule));
}
/* All outstanding data in existing flows has been accounted, so it's a
diff --git a/ofproto/ofproto-provider.h b/ofproto/ofproto-provider.h
index f2274ef..95bda33 100644
--- a/ofproto/ofproto-provider.h
+++ b/ofproto/ofproto-provider.h
@@ -71,6 +71,10 @@ struct ofproto {
struct oftable *tables;
int n_tables;
+ /* Optimisation for flow expiry.
+ * These flows should all be present in tables. */
+ struct list expirable; /* Expirable 'struct rule"s in all tables. */
+
/* OpenFlow connections. */
struct connmgr *connmgr;
@@ -221,6 +225,10 @@ struct rule {
enum nx_flow_monitor_flags monitor_flags;
uint64_t add_seqno; /* Sequence number when added. */
uint64_t modify_seqno; /* Sequence number when changed. */
+
+ /* Optimisation for flow expiry. */
+ struct list expirable; /* In ofproto's 'expirable' list if this rule
+ * is expirable, otherwise empty. */
};
static inline struct rule *
diff --git a/ofproto/ofproto.c b/ofproto/ofproto.c
index 9bae971..cd9d328 100644
--- a/ofproto/ofproto.c
+++ b/ofproto/ofproto.c
@@ -419,6 +419,7 @@ ofproto_create(const char *datapath_name, const char
*datapath_type,
ofproto->max_ports = OFPP_MAX;
ofproto->tables = NULL;
ofproto->n_tables = 0;
+ list_init(&ofproto->expirable);
ofproto->connmgr = connmgr_create(ofproto, datapath_name, datapath_name);
ofproto->state = S_OPENFLOW;
list_init(&ofproto->pending);
@@ -3162,6 +3163,7 @@ add_flow(struct ofproto *ofproto, struct ofconn *ofconn,
rule->ofpacts_len = fm->ofpacts_len;
rule->evictable = true;
rule->eviction_group = NULL;
+ list_init(&rule->expirable);
rule->monitor_flags = 0;
rule->add_seqno = 0;
rule->modify_seqno = 0;
@@ -4824,6 +4826,9 @@ oftable_remove_rule(struct rule *rule)
classifier_remove(&table->cls, &rule->cr);
eviction_group_remove_rule(rule);
+ if (!list_is_empty(&rule->expirable)) {
+ list_remove(&rule->expirable);
+ }
}
/* Inserts 'rule' into its oftable. Removes any existing rule from 'rule''s
@@ -4835,6 +4840,11 @@ oftable_replace_rule(struct rule *rule)
struct ofproto *ofproto = rule->ofproto;
struct oftable *table = &ofproto->tables[rule->table_id];
struct rule *victim;
+ bool may_expire = rule->hard_timeout || rule->idle_timeout;
+
+ if (may_expire) {
+ list_insert(&ofproto->expirable, &rule->expirable);
+ }
victim = rule_from_cls_rule(classifier_replace(&table->cls, &rule->cr));
if (victim) {
--
1.7.10.4
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
