This was a class assignment in the 15-213 class at Carnegie Mellon. The staff had set up a test suite and an online leaderboard to rank the speed of each student's malloc implementation.
I figured out that the test cases allocated a disproportionate amount of X-byte blocks. I was able to get to the top by hardcoding a specific freelist just for X-byte blocks.
Learned a lesson about easily it is to game a benchmark :)
This is a useful lesson! Tuning an allocator for a single application reduces to optimizing for that application’s empirical allocation patterns (sizes, lifetimes, access, usage).
Implemented my own specialized memory allocator 26+ yrs ago. (Y2K timeframe) Probably older than the most of the CMU students :-(
Use pre-allocated pools with array of indexes, free/allocation idx for alloc and free.
Con: Fixed pool size and fixed amount of memory can be allocated per pool.
Pro: constant cost operations per alloc/free via Atomic inc/dec of idx - no linklist tranversing ; Can be alloc in kernel space and free in user space (linux/QNX) and in multiple user processes when memory pools are in shmem; Run very will in SMP environment without any locks - all memory contentions were handled with atomic +/- alloc/free idx.
Same source code run in QNX, vxworks and linux (kernel and user space) at that time.
I was trying to get my c wasm module down in size and emmalloc is pretty small, but it requires a lot of js glue to make it work, but using a small allocator like walloc requires no glue and it's insanely small. I got my module down from 27kb to 17kb.
I'm gonna read this article and try making my own allocator next.
We also did this at Harvard in CS 61, which is the Systems Programming and Machine Organization class currently taught by Eddie Kohler (and has been for a while). It's a good exercise :)
Text appears generated. Minus 10 points.
This was a class assignment in the 15-213 class at Carnegie Mellon. The staff had set up a test suite and an online leaderboard to rank the speed of each student's malloc implementation.
I figured out that the test cases allocated a disproportionate amount of X-byte blocks. I was able to get to the top by hardcoding a specific freelist just for X-byte blocks.
Learned a lesson about easily it is to game a benchmark :)
This is a useful lesson! Tuning an allocator for a single application reduces to optimizing for that application’s empirical allocation patterns (sizes, lifetimes, access, usage).
Implemented my own specialized memory allocator 26+ yrs ago. (Y2K timeframe) Probably older than the most of the CMU students :-(
Use pre-allocated pools with array of indexes, free/allocation idx for alloc and free.
Con: Fixed pool size and fixed amount of memory can be allocated per pool.
Pro: constant cost operations per alloc/free via Atomic inc/dec of idx - no linklist tranversing ; Can be alloc in kernel space and free in user space (linux/QNX) and in multiple user processes when memory pools are in shmem; Run very will in SMP environment without any locks - all memory contentions were handled with atomic +/- alloc/free idx.
Same source code run in QNX, vxworks and linux (kernel and user space) at that time.
I was trying to get my c wasm module down in size and emmalloc is pretty small, but it requires a lot of js glue to make it work, but using a small allocator like walloc requires no glue and it's insanely small. I got my module down from 27kb to 17kb.
I'm gonna read this article and try making my own allocator next.
We also did this at Harvard in CS 61, which is the Systems Programming and Machine Organization class currently taught by Eddie Kohler (and has been for a while). It's a good exercise :)
[dead]