Shibboleth SP getting progressively slower
cantor.2 at osu.edu
Thu Jan 19 18:09:08 GMT 2012
On 1/19/12 10:22 AM, "Steve Thorpe" <thorpe at mailbox.mcnc.org> wrote:
>It seems likely we also encountered this bottleneck at MCNC, though it
>wasn't with load testing per se.
It is the same issue, and it does cause this problem. I do expire the
reverse index at fault here, but the problem is it expires as a unit, and
the expiration slides forward every time a new session shows up with the
same NameID. This would be exactly such a case, so it never purges that
structure and it gets slower and slower. No doubt that's what the OP is
Thanks for highlighting this, I hadn't identified that particular vector.
Fixing it is hard, because of the goal of keeping it all in memory but
also supporting back channel logout (or even front channel with no cookie).
I think given that the load testing case is artificial, and that
monitoring can usually be done without using a fixed NameID, though, it's
still relatively minor in the scheme of things.
More information about the users