FYI: updated guidance on heap size for Shibboleth IdP deployments
I wanted to highlight for the Shibboleth community that we have updated the guidance we give on how much memory to allocate to the Java heap when running the V2 or V3 IdP software.
Historically, we suggested a heap size of 512M as a safe guess, but people have been experiencing problems with refresh of larger metadata files (such as those in the UK Access Mgmt Federation or InCommon) and the evidence and testing suggest that lack of memory is the root cause of a lot of those problems.
Because of bugs , the out-of-memory condition is getting logged on DEBUG instead of ERROR, so it's been masked up to this point for a lot of people. The bug entries below also include information on working around the bugs for now by raising a specific logging class to DEBUG to make sure these problems are actually visible if you do run into them.
After some testing, we have raised the guidance on the various pages where we documented the various supported containers from 512M to 1024M as a suggested heap size. This is particularly relevant for larger metadata files in the 25-30M range, and that in turn is going to become more common with the incorporation of eduGAIN metadata into more local HE federations.
Of course, those running with much more limited metadata sources are not likely to need anything close to that much memory.
-- Scott, on behalf of the Shibboleth Project team