Scott,
For someone who doesn't use LoadRunner the term "vuser" is foreign. I
don't speak for Andrew, but I don't use LoadRunner.
A jMeter thread runs a script. By default it runs without pauses or
without "think time." Most people would extrapolate that 20 jMeter
threads translate to magnitude of times more real users. What's the
magnitude? I have no idea. After all, the purpose of the tests
discussed in this thread was not for capacity planning purposes but to
prove that no errors occurred in the distributed TicketRegistry when
the servers were subjected to heavy load. The CPU utilization of my
test servers was near 100% and the load average on them was over 2.
As they say, "your mileage will vary." ;-)
Adam
Scott Battaglia wrote:
On Wed, Oct 29, 2008 at 10:13 AM, Andrew
Ralph Feller, afelle1 <[EMAIL PROTECTED]> wrote:
Scott,
Jmeter threads simulate a run of a stress test. In the event of the
Jmeter test Adam is referring to, which I posted some time ago,
consists of logging into CAS, requesting a CAS protected resource, and
logging out of CAS.
Thanks, but it still doesn't correlate JMeter to LoadRunner. Does it
mean you were running only the equivalent of 20 vusers (a LoadRunner
term)?
-Scott
A-
Interesting, thanks! When I get a moment,
we'll probably run our load testing with the same configuration option
to see if there is a performance difference (we didn't see your initial
problem in our testing, but for safety sake we should try out this new
method). We used LoadRunner so its tough for me to see how your test
results correlate to our test results (is a JMeter thread the same as a
LoadRunner VUser?).
-Scott
-Scott Battaglia
PGP Public Key Id: 0x383733AA
LinkedIn: http://www.linkedin.com/in/scottbattaglia
On Tue, Oct 28, 2008 at 1:41 PM, Adam Rybicki <[EMAIL PROTECTED]> wrote:
Scott,
No noticeable performance difference. The numbers I had before had too
many errors to use them for comparison. My test environment uses a
dummy authentication handler because real authentication wasn't a goal
of this test. The jMeter script requests the CAS login page, extracts
the hidden lt field, and submits authentication. After verifying that
it received a TGT, the script performs the following steps:
- Access a protected resource (service)
- Redirect to CAS
- CAS redirects with a service ticket
- Service validates the ticket
- Service displays the protected
resource
- Script verifies that a message from
protected resource is present.
These steps are "real" in
contrast to authentication. The TGT has to be validated, ST has to be
issued, and the ST needs to be validated. Under a load of 20 jMeter
threads the above steps complete in 300 milliseconds on average, which
in my opinion is very fast. Under the light load of 1 jMeter thread
these steps take about 20 milliseconds as before. The memcached server
and its replica run on the same servers as CAS.
For your information, I also performed a test with JpaTicketRegistry. I
used MySQL running on one of the same servers as CAS. The same steps
listed above complete in 340 milliseconds using 20 jMeter threads. This
is a small impact considering that all the tickets go through the
database.
My servers were not evenly loaded in these tests because I only needed
one Apache for load-balancing and SSL for CAS. I also used one MySQL
on one of the servers for testing JpaTicketRegistry. Since I didn't
have the recently released Apache 2.2.10, I could not take advantage of
its "smart" load-balancer option "bybusyness" to get optimal
performance. I manually adjusted the lbfactor option until I got the
best response time.
Fascinating stuff.
Adam
Scott Battaglia wrote:
On Mon, Oct 27, 2008 at 3:19
PM, Adam Rybicki <[EMAIL PROTECTED]>
wrote:
Scott,
Great fix! I cannot make it fail again.
Any noticeable performance difference?
-Scott
Thanks,
Adam
Scott
Battaglia wrote:
Let me know how the test goes. Also if you find any bugs since again,
it was written in about 30 seconds :-)
-Scott
-Scott Battaglia
PGP Public Key Id: 0x383733AA
LinkedIn: http://www.linkedin.com/in/scottbattaglia
On Mon, Oct 27, 2008 at 1:26 PM, Adam Rybicki <[EMAIL PROTECTED]> wrote:
Scott,
Great. I will grab it and retest. This will probably solve the issue.
Adam
Scott
Battaglia wrote:
On Fri, Oct 24, 2008 at 7:58 PM, Adam Rybicki <[EMAIL PROTECTED]> wrote:
Scott,
I mis-diagnosed the issue. I just ran the same test, except I only ran
one instance of memcached. I am getting a high error rate on ticket
validations. So, it has nothing to do with memcached replication. To
investigate further, I disabled the second CAS server, and all errors
are gone. Of course that is not a viable workaround. :-)
My guess is that the error occurs when a ticket issued by one CAS
server is being validated on another CAS server. I could not find a
way to enable debug logging in /cas/serviceValidate, but I think I have
found a major clue. It took most of the day today to hunt this down.
With a single instance of memcached running in verbose mode you can see
a sequence of messages like this:
<11 add
ST-8023-M0sU2U2ijyQ53QPYWnGm-arybicki1 1 300 2689
>11 STORED
<7 get ST-8023-M0sU2U2ijyQ53QPYWnGm-arybicki1
>7 sending key ST-8023-M0sU2U2ijyQ53QPYWnGm-arybicki1
>7 END
<7 replace ST-8023-M0sU2U2ijyQ53QPYWnGm-arybicki1 1 300 2689
>7 STORED
<7 delete ST-8023-M0sU2U2ijyQ53QPYWnGm-arybicki1 0
>7 DELETED
This is when everything
went OK. The sequence below, however, represents a service ticket that
failed to validate. That's apparently because an attempt to read the
ticket was made before it was actually stored in cache!
<11 add
ST-8024-tKeeo5gYhjqoQzstAgqO-arybicki1 1 300 2689
<7 get ST-8024-tKeeo5gYhjqoQzstAgqO-arybicki1
>7 END
>11 STORED
There may be some code
that synchronizes access to the same object from the same client.
However, it would seem that the service ticket is returned by CAS
before it's actually stored in memcached. If this service ticket is
then presented to another instance of CAS for validation, it fails to
retrieve it from memcached because the "add" operation has not
completed.
Again, I have to emphasize that this is not an unrealistic test. The
jMeter is simply following redirects at the time of the failure, as a
browser would.
We never saw that in production and we ran 500 virtual users. However,
if you are experiencing it, you most likely could update the
MemcacheTicketRegistry to block on the Futures. I've actually updated
the code in HEAD with an option to block on Futures. :-)
I have not tried it at all, since I wrote it all of 30 seconds ago. You
can grab it from HEAD and try it out. The new property to enable it is
"synchronizeUpdatesToRegistry"
Let me know if it helps/doesn't help.
-Scott
Adam
Scott Battaglia wrote:
You have no need for sticky sessions. If you have two repcached
servers and you've told your CAS instance about both of them, the
memcached client essentially sees them as two memcached servers (since
its not familiar with repcached).
The memcached client works in that it takes a hash of the key and that
determines what instance of memcached/repcached to store the item on.
repcached will then do its async replication. When you come to
validate a ticket the memcached client will again hash the key to
determine what server the item is stored on. If that server is
unreachable (as determined by the memcached client) then it will try
the next likely server that would hold the data.
-Scott
-Scott Battaglia
PGP Public Key Id: 0x383733AA
LinkedIn: http://www.linkedin.com/in/scottbattaglia
On Fri, Oct 24, 2008 at 8:21 AM, Andrew Ralph Feller, afelle1 < [EMAIL PROTECTED]>
wrote:
So what you are saying is that even with replication enabled,
asynchronous replication CAS clusters should have sticky sessions on
regardless? I realize that synchronous replication CAS cluster have no
need of sticky sessions seeing as how it goes to all servers before the
user can move on.
Andrew
On 10/23/08 9:29 PM, "Scott Battaglia" < [EMAIL PROTECTED] < http://[EMAIL PROTECTED]> > wrote:
It actually shouldn't matter if the async works or not. The memcache
clients are designed to hash to a particular server and only check the
backup servers if the primary isn't available.
So you should always be validating against the original server unless
its no longer there.
-Scott
-Scott Battaglia
PGP Public Key Id: 0x383733AA
LinkedIn: http://www.linkedin.com/in/scottbattaglia
I have run into a issue with MemCacheTicketRegistry and was wondering
if you have any thoughts. I didn't want to create a new thread for
this note. Anyone else with comments should feel free to reply, too.
;-)
My tests have shown that when a ticket is generated on a CAS cluster
member it may sometimes fail to validate. This is apparently because
the memcached asynchronous replication did not manage to send the
ticket replica in time. Fast as repcached may be, under a relatively
light load, ST validation failed in 0.1% of the cases, or once in 1000
attempts. It would seem that the following tasks should be fairly
complex:
- Browser accesses a CAS-protected
service
- Service redirects to CAS for
authentication
- CAS validates the TGT
- CAS issues the ST for service
- CAS redirects the browser to service
- Service sends the ST for validation
-
But they are fast!
My jMeter testing showed this taking 28 milliseconds under light load
on CAS server , which is amazingly fast. Please note that in real
life, this can be just as fast because the browser, CAS, and service
perform these steps without the user slowing them down. CAS is indeed
a lightweight system, and memcached does nothing to slow it down. It
seems that in 0.1% of the cases this outperforms repcached under light
load. The bad news is that under heavy load the failure rate
increases. I've seen as bad as 8% failure rate.
Have you or anyone else seen this? Have you had to work around this?
Thanks,
Adam
Scott Battaglia
wrote:
On Tue, Oct 14, 2008 at 11:15 AM, Andrew Ralph Feller, afelle1 < [EMAIL PROTECTED]
< http://[EMAIL PROTECTED]> > wrote:
Hey Scott,
Thanks for answering some questions; really appreciate it. Just a
handful more:
- What happens whenever the
server it intends to replicate with is down?
-
It doesn't replicate :-) The client will send its request to the
primary server and if the primary server is down it will replicate to
the secondary. The repcache server itself will not replicate to the
other server if it can't find it.
-
-
-
- What happens whenever it
comes back up?
-
The repcache servers will sync with each other. The memcache clients
will continue to function as they should
-
-
-
- Does the newly recovered
machine synchronize itself with the other servers?
-
The newly recovered machine with synchronize with its paired memcache
server.
-Scott
-
-
-
Thanks,
Andrew
On Tue, Oct 14, 2008 at 10:38 AM, Andrew Ralph
Feller, afelle1 < [EMAIL PROTECTED] < http://[EMAIL PROTECTED]>
< http://[EMAIL PROTECTED]> > wrote:
Scott,
I've looked at the sample configuration file on the JA-SIG wiki,
however I was curious how memcached handles cluster membership for lack
of a better word. One of the things we are getting burned on by
JBoss/Jgroups is the frequency the cluster is being fragmented.
Thanks,
Andrew
On Mon, Oct 13, 2008 at 11:41 PM, Patrick
Hennessy < [EMAIL PROTECTED] < http://[EMAIL PROTECTED]>
< http://[EMAIL PROTECTED]> < http://[EMAIL PROTECTED]>
> wrote:
I've been working on updating from 3.2 to 3.3 and wanted to give
memcached a try instead of JBoss. I read Scott's message about
performance and we've had good success here with memcached for other
applications. It also looks like using memcached instead of JBoss will
simplify the configuration changes for the CAS server.
I do have the JBoss replication working with CAS 3.2 but pounding the
heck out of it with JMeter will cause some not so nice stuff to happen.
I'm using VMWare VI3 and configured an isolated switch for the
clustering and Linux-HA traffic. I do see higher traffic levels coming
to my cluster in the future, but I'm not sure if they'll meet the levels
from my JMeter test. (I'm just throwing this out there because of the
recent Best practice thread.)
If I use memcached, is the ticketRegistryCleaner not needed anymore? I
left those beans in the ticketRegistry.xml file and saw all kinds of
errors. After taking it out it seems to load fine and appears to work,
but I wasn't sure what the behavior is and I haven't tested it further.
What if memcached fills up all the way? Does anyone have a general
idea of how much memory to allocate to memcached with regards to
concurrent logins and tickets stored?
Thanks,
Pat
--
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Senior Systems Specialist
Division of Information and Educational Technology
Delaware Technical and Community College
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
_______________________________________________
Yale CAS mailing list
[email protected] <http://[email protected]>
<http://[email protected]> <http://[email protected]>
_______________________________________________
Yale CAS mailing list
[email protected] <http://[email protected]>
<http://[email protected]> <http://[email protected]>
--
Andrew R. Feller, Analyst
Information Technology Services
200 Fred Frey Building
Louisiana State University
Baton Rouge, LA 70803
(225) 578-3737 (Office)
(225) 578-6400 (Fax)
_______________________________________________
Yale CAS mailing list
[email protected] <http://[email protected]>
_______________________________________________
Yale CAS mailing list
[email protected]
http://tp.its.yale.edu/mailman/listinfo/cas
_______________________________________________
Yale CAS mailing list
[email protected]
http://tp.its.yale.edu/mailman/listinfo/cas
|
smime.p7s