From time to time tcmalloc acquires memory from the system using mmap, and returns it to the system using madvise. These can be expensive operations, and they hold a lock so they can create a bottleneck. There aren't adequate metrics to diagnose issues related to this, and to assess the effect of making changes such as enabling TCMALLOC_AGGRESSIVE_DECOMMIT. The following could be helpful:

• number of mmap operations to acquire memory
• number of bytes acquired
• cumulative time spent in such operations
• number of madvise oeprations to release memory
• number of bytes released
• cumulative time spent in such operations
• number of bytes allocated using decommitted memory (which kernel then must re-commit)