We have a product which spawns some tens of thousands of processes in the course of even a short run, and the 'converting tb5 to MDB' stage uses a couple of gigabytes of memory and takes tens of minutes.
While it's doing this, I ran out of curiosity another copy of vtune and used it to sample the first one. The hottest spot is aStringAlloc in libvtsa_gen.so; 99% of the time is being spent at the single 'leal (%edi,%eax,1),%edx' instruction which is used to check whether some entry in a list has a small enough associated value.
I've ordered four gigabytes more memory, which ought to be enough to have the job complete, but I'm wondering whether this is a pathological case for which VTune is not well optimised.
