Back in the 1980's I co-wrote a computer algebra system, that ran for example on many colleagues' first computer: The first Macintosh. Resources were scarce, ev

Back in the 1980's I co-wrote a computer algebra system, that ran for example on many colleagues' first computer: The first Macintosh. Resources were scarce, everything mattered. Malloc attempts to work acceptably well under diverse scenarios. Specific programs such as ours can adopt simpler memory disciplines, supported by an interface with malloc. Almost all of our memory needs could be served by pools of fixed-size slugs of memory. Each pool experienced considerable churn and reuse, but the total memory needed by each pool was unimodal, growing to a peak then subsiding until the pool was no longer needed. We chose a universal fixed-size block to be served by malloc. We kept these blocks on a master stack, to assign to pools as needed. We kept the slugs in each pool on stacks, to reuse as needed. Closing a pool returned its blocks to the master stack. There's nothing novel here; this is about as dumb a protocol as one can imagine, and correspondingly faster than anything else we tried. I still believe the lesson today: One's core algorithms need to be crafted to cooperate with the fastest memory protocols. The doctor can't do everything without the patient's cooperation. For example, I believe that the approach taken by Rust will ascend as we better learn how to work with it.