AccountsDb Async Fetching Service

[jon-chuang]

Problem

Currently, MMap access can trigger page fault. Typical usage is sequential load_with_fixed_root.

Proposed Solution

The solution is an AccountsDb async worker pool that context switches to interleave page faults.

The design is as follows: AccountsDb or Accounts exposes some methods that rely on async workers instead of sequentially performing IO.

TO DO: MORE DETAILS

Choices and Evidence

According to https://github.com/jimblandy/context-switch, a tokio context switch (similar to go green threads), is lower than kernel thread context switch/thread creation. (although apparently not for blocking IO 😢 )

An alternative is may, which implements something similar to coroutines: example use

While asking a threadpool to perform IO would actually increase the IO time in the non-page fauting case, this time is actually very small and doesn't add much to the critical path. But, it can drastically reduce the time waiting around for page faults, possible even on Optane-like low latency SSDs.