Tune flow table lookup in net/openvswitch, replacing a call to
the slow-but-safe memcmp() in lib/string.c with a key comparator
routine that presumes most comparisons will succeed. Besides
avoiding an early-exit test on each iteration, it also compares
keys 4 or 8 bytes at a time on architectures that can load an
unaligned long efficiently.
On a 3.2GHz Xeon (5679) this patch reduces the minimum back-to-back
hot-cache latency of a 128-byte key comparison by 7x, from 130ns with
the default byte-at-a-time memcmp() in lib/string.c down to 17ns.
More important, replacing the default memcmp() with this specialized
routine speeds up openvswitch's packet rate by 10% in a closed-loop
benchmark that simply routes traffic from one tap device to another.
Signed-off-by: Peter Klausler <[email protected]>
---
net/openvswitch/flow.c | 33 ++++++++++++++++++++++++++++++++-
1 file changed, 32 insertions(+), 1 deletion(-)
diff --git a/net/openvswitch/flow.c b/net/openvswitch/flow.c
index 67a2b78..d5facf6 100644
--- a/net/openvswitch/flow.c
+++ b/net/openvswitch/flow.c
@@ -764,6 +764,37 @@ u32 ovs_flow_hash(const struct sw_flow_key *key, int key_len)
return jhash2((u32 *)key, DIV_ROUND_UP(key_len, sizeof(u32)), 0);
}
+/*
+ * Key comparison routine, optimized for the common case of
+ * equality due to low average hash collision frequency
+ * (1.5 mean items per nonempty bucket when total table item
+ * count equals the number of buckets, which is when openvswitch
+ * expands its hash table).
+ */
+static bool equal_keys(const struct sw_flow_key *key1,
+ const struct sw_flow_key *key2,
+ size_t key_len)
+{
+ const char *cp1 = (const char *)key1;
+ const char *cp2 = (const char *)key2;
+ long diffs = 0;
+
+#ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
+ {
+ const long *lp1 = (const long *)cp1;
+ const long *lp2 = (const long *)cp2;
+ for (; key_len >= sizeof(long); key_len -= sizeof(long))
+ diffs |= *lp1++ ^ *lp2++;
+ cp1 = (const char *)lp1;
+ cp2 = (const char *)lp2;
+ }
+#endif
+
+ while (key_len-- > 0)
+ diffs |= *cp1++ ^ *cp2++;
+ return diffs == 0;
+}
+
struct sw_flow *ovs_flow_tbl_lookup(struct flow_table *table,
struct sw_flow_key *key, int key_len)
{
@@ -777,7 +808,7 @@ struct sw_flow *ovs_flow_tbl_lookup(struct flow_table *table,
hlist_for_each_entry_rcu(flow, head, hash_node[table->node_ver]) {
if (flow->hash == hash &&
- !memcmp(&flow->key, key, key_len)) {
+ equal_keys(&flow->key, key, key_len)) {
return flow;
}
}
--
1.8.2.1
On Fri, 2013-04-26 at 17:46 -0400, Peter Klausler wrote:
> Tune flow table lookup in net/openvswitch, replacing a call to
> the slow-but-safe memcmp() in lib/string.c with a key comparator
> routine that presumes most comparisons will succeed. Besides
> avoiding an early-exit test on each iteration, it also compares
> keys 4 or 8 bytes at a time on architectures that can load an
> unaligned long efficiently.
>
> On a 3.2GHz Xeon (5679) this patch reduces the minimum back-to-back
> hot-cache latency of a 128-byte key comparison by 7x, from 130ns with
> the default byte-at-a-time memcmp() in lib/string.c down to 17ns.
>
> More important, replacing the default memcmp() with this specialized
> routine speeds up openvswitch's packet rate by 10% in a closed-loop
> benchmark that simply routes traffic from one tap device to another.
>
> Signed-off-by: Peter Klausler <[email protected]>
> ---
> net/openvswitch/flow.c | 33 ++++++++++++++++++++++++++++++++-
> 1 file changed, 32 insertions(+), 1 deletion(-)
>
> diff --git a/net/openvswitch/flow.c b/net/openvswitch/flow.c
> index 67a2b78..d5facf6 100644
> --- a/net/openvswitch/flow.c
> +++ b/net/openvswitch/flow.c
> @@ -764,6 +764,37 @@ u32 ovs_flow_hash(const struct sw_flow_key *key, int key_len)
> return jhash2((u32 *)key, DIV_ROUND_UP(key_len, sizeof(u32)), 0);
> }
>
> +/*
> + * Key comparison routine, optimized for the common case of
> + * equality due to low average hash collision frequency
> + * (1.5 mean items per nonempty bucket when total table item
> + * count equals the number of buckets, which is when openvswitch
> + * expands its hash table).
> + */
> +static bool equal_keys(const struct sw_flow_key *key1,
> + const struct sw_flow_key *key2,
> + size_t key_len)
> +{
> + const char *cp1 = (const char *)key1;
> + const char *cp2 = (const char *)key2;
> + long diffs = 0;
> +
> +#ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
> + {
> + const long *lp1 = (const long *)cp1;
> + const long *lp2 = (const long *)cp2;
> + for (; key_len >= sizeof(long); key_len -= sizeof(long))
> + diffs |= *lp1++ ^ *lp2++;
> + cp1 = (const char *)lp1;
> + cp2 = (const char *)lp2;
> + }
> +#endif
This seems a suboptimal condition, a bit tweaked for x86
Some 32bit arches do not have CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS,
still struct sw_flow_key is 32bit word aligned.
Either its a generic bcmp(s1,s2,len) function without any information on
s1/s2 alignment, and it should not a private ovs thing,
or
its OVS private helper, and you can make sure alignof(struct
sw_flow_key) == alignof(unsigned long) and remove the #ifdef
CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
Check include/net/flow.h for an example
struct flowi4 {
...
} __attribute__((__aligned__(BITS_PER_LONG/8)));