Hi Andy,
On Sun, 7 May 2006, Andy Gospodarek wrote:
On Sun, May 07, 2006 at 11:32:00PM +0800, Wensong Zhang wrote:
Hi Andy,
Yes, the original sychronziation design is a sort of arbitary or
compromised solution. We don't want to synchronize every state change from
master to backup load balancer, because we were afraid that there were too
much state change synchronization messages and there would be some
performance penalty. So, we only sychronize the connection of TCP
ESTABLISHED state or UDP to backup load balancer, and use the timeout of 3
minutes.
Your change is probably ok, but we should be aware that it may create a
lot of connection entries at backup load balancer for TCP applications,
which is much more than that at master load balancer, because there is no
connection close synchronization and timeout is changed to 3 minutes to 15
minutes.
Wensong,
This is an interesting point. Do you feel it would be reasonable to
synchronize states for extablished connections and terminating
connections rather than all state changes? It would seem keeping the
table at a minimum size would be desireable without tracking every state
change since that could create too much noice between master and backup
-- especially for short-lived connections.
Yeah, it's reasonable, but we need to synchronize updating states of
established connections too when established connections keep receiving
packets, so that we can maintain the timeout of those established
connections more accurately. It's really a trade-off to keep connection
timeout more accurate in the mean while keeping the volume of
synchronization messages as low as possible.
I could work some more to expand and test this patch to include more
detailed connection tracking if you (or anyone else) feel it would be
valuable.
Maybe it's better to write a more detailed design first.
Thanks,
Wensong
-andy
The simple solution is probably to make timeout value tuneable, so that
users can tune it for their application. The better solution is to
synchronize very connection state change from master to backup, so that
backup have almost the same state of connections.
Thanks,
Wensong
On Thu, 4 May 2006, Andy Gospodarek wrote:
Instead of using the default timeout of 3 minutes, this uses the timeout
specific to the protocol used for the connection. The 3 minute timeout
seems somewhat arbitrary (though I know it is used other places in the
ipvs code) and when failing over it would be much nicer to use one of
the configured timeout values.
Signed-off-by: Andy Gospodarek <[EMAIL PROTECTED]>
---
ip_vs_sync.c | 5 +++--
1 files changed, 3 insertions(+), 2 deletions(-)
diff --git a/net/ipv4/ipvs/ip_vs_sync.c b/net/ipv4/ipvs/ip_vs_sync.c
--- a/net/ipv4/ipvs/ip_vs_sync.c
+++ b/net/ipv4/ipvs/ip_vs_sync.c
@@ -67,7 +67,6 @@ struct ip_vs_sync_conn_options {
struct ip_vs_seq out_seq; /* outgoing seq. struct */
};
-#define IP_VS_SYNC_CONN_TIMEOUT (3*60*HZ)
#define SIMPLE_CONN_SIZE (sizeof(struct ip_vs_sync_conn))
#define FULL_CONN_SIZE \
(sizeof(struct ip_vs_sync_conn) + sizeof(struct ip_vs_sync_conn_options))
@@ -279,6 +278,7 @@ static void ip_vs_process_message(const
struct ip_vs_sync_conn *s;
struct ip_vs_sync_conn_options *opt;
struct ip_vs_conn *cp;
+ struct ip_vs_protocol *pp;
char *p;
int i;
@@ -337,7 +337,8 @@ static void ip_vs_process_message(const
p += SIMPLE_CONN_SIZE;
atomic_set(&cp->in_pkts, sysctl_ip_vs_sync_threshold[0]);
- cp->timeout = IP_VS_SYNC_CONN_TIMEOUT;
+ pp = ip_vs_proto_get(s->protocol);
+ cp->timeout = pp->timeout_table[cp->state];
ip_vs_conn_put(cp);
if (p > buffer+buflen) {
-
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
-
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
