5 ms·

Might one do a cache-aware code competition? Given the simulator's determinism and cache. As attention to cache-friendly data layout seems often lagging its im

Might one do a cache-aware code competition? Given the simulator's determinism and cache. As attention to cache-friendly data layout seems often lagging its importance ... err, and the world doesn't yet have enough premature optimization. Implement this algorithm; it will be challenged with <large n> input sets; here is a training set generator; your competition set runtimes will be aggregated with <metric>; minimize it. Does anything like this exist already?

zozbot234 4y ago
Cache-oblivious code is a lot more successful in the real world, because systems change over time. Hardwiring code to rely on a rigidly predefined amount of cache is the worst kind of premature optimization.
- binarycrusader 4y ago
  Isn’t “a lot more successful” is only true more generally? Many developers have fixed platform targets for the life of a product (e.g., consoles, embedded, microarch). So I’d agree in general but hesitate to say “a lot more successful”.
- mncharity 4y ago
  So the concern is overfitting to a single target? Hmm, then perhaps: here are several/parameterized cache models to develop on; one/several will be used to judge the competition? > Cache-oblivious code is a lot more successful [...] Hardwiring code to rely on a rigidly predefined amount of cache is the worst Yes, but ... Creating cache-oblivious code requires devs non-oblivious to cache. Perhaps a kata/competition might encourage such.
user062 4y ago
checkout the cache lab [1] [1] https://csapp.cs.cmu.edu/3e/labs.html https://csapp.cs.cmu.edu/3e/labs.html
- mncharity 4y ago
  > https://csapp.cs.cmu.edu/3e/labs.html https://csapp.cs.cmu.edu/3e/labs.html Accessing cachelab.tar there regrettably requires login. Also on https://csapp.cs.cmu.edu/public/labs.html https://csapp.cs.cmu.edu/public/labs.html . Though the other Cache Lab links work. But googling "cachelab" currently turns up similar elsewhere. And a few students' solutions. Tnx!
  - user062 4y ago
    You're welcome, the self-study handout [1] doesn't require login. [1] http://csapp.cs.cmu.edu/3e/cachelab-handout.tar http://csapp.cs.cmu.edu/3e/cachelab-handout.tar

zozbot234 4y ago

Cache-oblivious code is a lot more successful in the real world, because systems change over time. Hardwiring code to rely on a rigidly predefined amount of cache is the worst kind of premature optimization.

binarycrusader 4y ago
Isn’t “a lot more successful” is only true more generally? Many developers have fixed platform targets for the life of a product (e.g., consoles, embedded, microarch). So I’d agree in general but hesitate to say “a lot more successful”.
mncharity 4y ago
So the concern is overfitting to a single target? Hmm, then perhaps: here are several/parameterized cache models to develop on; one/several will be used to judge the competition? > Cache-oblivious code is a lot more successful [...] Hardwiring code to rely on a rigidly predefined amount of cache is the worst Yes, but ... Creating cache-oblivious code requires devs non-oblivious to cache. Perhaps a kata/competition might encourage such.

user062 4y ago

checkout the cache lab [1] [1] https://csapp.cs.cmu.edu/3e/labs.html https://csapp.cs.cmu.edu/3e/labs.html

mncharity 4y ago
> https://csapp.cs.cmu.edu/3e/labs.html https://csapp.cs.cmu.edu/3e/labs.html Accessing cachelab.tar there regrettably requires login. Also on https://csapp.cs.cmu.edu/public/labs.html https://csapp.cs.cmu.edu/public/labs.html . Though the other Cache Lab links work. But googling "cachelab" currently turns up similar elsewhere. And a few students' solutions. Tnx!
- user062 4y ago
  You're welcome, the self-study handout [1] doesn't require login. [1] http://csapp.cs.cmu.edu/3e/cachelab-handout.tar http://csapp.cs.cmu.edu/3e/cachelab-handout.tar