Comments inline..
Gianny Damour wrote:
On 17/02/2007, at 3:32 AM, Dave Colasurdo wrote:
<snip>
Tomcat and Mod-jk provide the ability to have the load balancer lock a
particular httpsession (identified by session cookie) to a specific
cluster member. This cluster member will serve all the requests for
this session until the cluster member tops/fails and failover to a new
cluster member occurs. This "stickiness" is quite useful for the
case of multi-frame JSPs where it is important that all concurrent
requests are executing in the same JVM (as per Java EE
specifications). Hmmm, now that Terracotta provides the concept of a
global cross-node JVM, I'm wondering how important this is now ..
though still seems there would be performance overhead in locking
across JVMs.. so I think Sticky Session is still useful..
Hi,
I think that if you have sticky load balancing, you do not need to
acquire a distributed lock to service concurrent requests: requests are
proxied to the node hosting the session state; each request acquires a
*local* read lock for the associated session (a read lock is acquire
such that requests bound to the same session can proceed concurrently).
This is only when a request is received by a node where the session
state cannot be found that a distributed lock is to be acquired.
Thanks,
Gianny
Hi Gianny,
Thanks for the input. I agree no distributed lock is required with
sticky session since all request are guaranteed to be running in the
same JVM.
I believe that Terracotta transparently shuttles the session data
between the cluster members without the application server ever knowing
that the session was actually created in a different JVM.
For the case where sticky session isn't configured it is possible for a
multiframe JSP to spray requests to multiple servers concurrently. The
servlet specification states:
"Within an application marked as distributable, all requests that are
part of a session must be handled by one Java Virtual Machine1 (“JVM”)
at a time."
So, for this case there would need to be some sort of distributed lock.
I believe Open Terracotta provides a mechanism to lock across JVMs
though believe sticky port would be a much better performing solution.
-Dave-
