[ovs-dev] [PATCH 3/4] ofproto-dpif: Push statistics less frequently.

Ethan Jackson ethan at nicira.com
Mon Apr 1 01:22:58 UTC 2013


The most natural place to push facet statistics is in
update_stats() where they're pulled from the datapath.  However,
under load, update_stats() can be called as many as 10 times per
second causing us to push statistics so frequently it hurts
performance.  By pushing statistics much less frequently, this
patch generates a roughly 8% improvement in TCP_CRR performance.

Signed-off-by: Ethan Jackson <ethan at nicira.com>
---
 ofproto/ofproto-dpif.c |   12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

diff --git a/ofproto/ofproto-dpif.c b/ofproto/ofproto-dpif.c
index ad78b52..2ac9434 100644
--- a/ofproto/ofproto-dpif.c
+++ b/ofproto/ofproto-dpif.c
@@ -909,6 +909,7 @@ lookup_ofproto_dpif_by_port_name(const char *name)
 static int
 type_run(const char *type)
 {
+    static long long int push_timer = LLONG_MIN;
     struct dpif_backer *backer;
     char *devname;
     int error;
@@ -922,6 +923,16 @@ type_run(const char *type)
 
     dpif_run(backer->dpif);
 
+    /* The most natural place to push facet statistics is when they're pulled
+     * from the datapath.  However, when there are many flows in the datapath,
+     * this expensive operation can occur so frequently, that it reduces our
+     * ability to quickly set up flows.  To reduce the cost, we push statistics
+     * here instead. */
+    if (time_msec() > push_timer) {
+        push_timer = time_msec() + 2000;
+        push_all_stats();
+    }
+
     if (backer->need_revalidate
         || !tag_set_is_empty(&backer->revalidate_set)) {
         struct tag_set revalidate_set = backer->revalidate_set;
@@ -4211,7 +4222,6 @@ update_subfacet_stats(struct subfacet *subfacet,
         facet_account(facet);
         facet->accounted_bytes = facet->byte_count;
     }
-    facet_push_stats(facet);
 }
 
 /* 'key' with length 'key_len' bytes is a flow in 'dpif' that we know nothing
-- 
1.7.9.5




More information about the dev mailing list