Potential perf improvements

[matklad]

0. Look at what APIs jemalloc itself has for introspection
1. Replace pub static TID: RefCell<usize> = RefCell::new(0); and the like with pub static TID: Cell<usize> = Cell::new(0);
2. Replace with const thread locals once it is stable Tracking Issue for const-initialized thread locals rust-lang/rust#84223
3. Replace with thread::current().id().as_u64() once that is stable
4. In general, minimize usage of thread-locals, they are quite slow in today's Rust
5. Spread out MEM_SIZE array more to avoid false sharing. Threads get sequential IDs, so may hammer the same cache line during allocation
6. Replace SeqCst with relaxed, I think that should be enough for our use case.
7. Remove TOTAL_MEMORY_USAGE as it's the sum of per-thread usages.
8. rand::thread_rng().gen_range(0, 100) feels like something optimizable: thread_rng is crypto secure I thing (or at least stroger than xorshift).
9. https://mirrors.edge.kernel.org/pub/linux/kernel/people/paulmck/perfbook/perfbook.html I think uses "counting memory usage by threads" as one of the running examples, might be worth reading (warning: deep rabbit hole).
10. https://blog.mozilla.org/nnethercote/category/memory-allocation/ might also be an interesting read

[pmnoxx]

@matklad Thanks. I added a benchmark, which shows performance degradation: #27

[pmnoxx]

7. Remove TOTAL_MEMORY_USAGE as it's the sum of per-thread usages.

@matklad I did some bench-marking, it saves round 0.4-0.6ns from each run.
#52

Though, this benchmark is single threaded. This variable could potentially, affect multi threaded applications more during high congestion.

[pmnoxx]

8. rand::thread_rng().gen_range(0, 100) feels like something optimizable: thread_rng is crypto secure I thing (or at least stroger than xorshift).

Yes, that's a big issue. I reduced time from 38ns to 12ns, by removing usage of rand. #53

[pmnoxx]

6. Replace SeqCst with relaxed, I think that should be enough for our use case.

I measured, there is no performance between using ::Relaxed vs ::SeqCst.
See https://stackoverflow.com/questions/53805142/why-memory-order-relaxed-performance-is-the-same-as-memory-order-seq-cst/53805377