11:03 on Intel platforms there are also specific instructions and mechanisms for cache interactions, such as prefetch, demote, and bypassing cache. Theres’s also CAT - “cache allocation technology” that can be used to design what the caches can or cannot hold
Timur, about double buffering and SeqLock, there is other way. I work in HFT, and we use it for some monitoring of big values. It is called Chen Algorithm, or Tripple buffering. It makes writer wait free, and read wait free. If there is only one reader, it is very simple code. With more readers it gets more complicated (and writer overheads scales with number of readers, but with one reader it is easy), but also doable. For multiple readers, we just take a lock, so only one reader actually reads.
Is the idea that you have one buffer for the writer and then two buffers per reader? So if you have one writer thread and three reader threads, you can use seven buffers for a wait-free sync?
48:00 there is some name shadowing for `new_coeffs` within `update_coeffs()`. I believe all but one on the first line should be something else, e.g. `new_storage`. Otherwise, thank you for this excellent talk!
For HFT, ideally, there'd be no OS at all. For a space shuttle OS with "hard real time" requirements, it's less about getting the lowest latency and more about making sure the OS is always running predictably at the worst case. So if you have a theoretical main loop that calls "OS_UpdateState(os_state)" to update the entire OS state on each loop, what you want is for that call's timings to be consistent regardless of what data you give it or what resources you create/delete that tick.
Yes, this find() will just get head->next. Further, I think _front suffix can go from pop/push because it is implicit from the thing being a stack. Also push_front is making an extra copy as it takes input by value.
Yes. I believe this code only supports a single updater thread (the fact that storage isn't atomic/synchronized in any way is an indication). I guess if there are multiple update threads and all of them aren't "hot", they can synchronize amongst themselves using a mutex
Yes, this can only work with a single producer because `storage` is not atomic. I also noticed a missing `)` of the for loop in `update_coeffs()`, which makes it a bit confusing to read.