note linux threading overhead blog

src/notes/reading_notes/index.md

@@ -1,5 +1,15 @@
 # Unstructured Reading Notes
 
+- [Measuring context switching and memory overheads for Linux
+    threads](https://eli.thegreenplace.net/2018/measuring-context-switching-and-memory-overheads-for-linux-threads/)
+    - Native POSIX Thread Library (NPTL) & futex several times faster than
+        old Linux thread
+    - bearable latency: on his i7-4771 in 2018,
+        multicore context switching took \~1.5µs, same core switching \~2.5µs,
+        launching thread \~5µs, launching process \~22.5µs\
+        comparison: memcpy 64 KiB took 3µs, Goroutine switching took 170ns
+    - lazy memory: 8 MiB virtual via `ulimit` (VM), \~8 KiB resident (RSS)
+        without touching stack
 - [Optimal Protocols for Studying & Learning,
     Andrew Huberman](https://www.youtube.com/watch?v=ddq8JIMhz7c)
     - neuroplasticity apply during fast eye movement sleep
@@ -11,8 +21,7 @@
         - do not matter for memory if wrong
     - perceived familiarity hinder memory
     - gap effect: pause in talk greatly reinforce memory
-- [ChatGPT from Scratch:
-    How to
+- [ChatGPT from Scratch: How to
     Train an Enterprise AI Assistant • Phil Winder • GOTO
     2023](https://www.youtube.com/watch?v=N53Gsz0Gm4c)
     - post-GPT are decoder-only