> As an example of the benefits of this approach, one service increased its time in TCMalloc from 2.7% to 3.5%, an apparent regression, but reaped improvements of 3.4% more requests-per-second, a 1.7% latency reduction, and a 6.5% reduction in peak memory usage!

This is the stuff of performance nightmares. Anyone thinking about optimization often will get single-tracked into the performance regression there and maybe not necessarily see the improved overall performance (requests per second).

> In Google’s data centers, this improvement reduced TLB stalls by 6% and memory fragmentation by 26%.

Yet after Ctr+F-ing the paper for the term, I have yet to find an accurate definition of "fragmentation". Keeping in mind that fragmentation is an allocator's major enemy, it bugs me to realize that there is no universally agreed-upon formulation yet.

Does anyone more knowledgeable have a more informed opinion on the matter?

Seems like a bit of a tragedy of the commons. Individual containers benefit from a faster malloc, but the whole fleet benefits from one doing more work.