TODO list for adaptive-hashmap hotcold benchmark

[rrnewton]

I've been fixing various things about Bench1.hs, but there's still a ways to go.

• [x] Need to finish refactoring to use GenericImpl
• [x] Need to make sure the client is allocating, even during read phase, for realism.
• [x] Need to use random keys rather than consecutive
  • [x] Computing these keys inline rather than precomputing will help with creating a nontrivial phase2 client workload.
• [x] Need to introduce a keyspace size parameter and get rid of the mod hack that's in there ("Temp/HACK modulus").
• [x] Related to previous, it's good for the client to actual use values retrieved via get, for example summing them.
• [x] Probably should make AdaptiveMap use INLINE rather than INLINABLE, because it helps to be certain there.
• [x] We have two copies of CTrie in there... delete one.

Remaining mysteries/questions:

• [x] Adaptive does reasonably well (close to pure) for coldPhase, but why doesn't c-adaptive yet do better?

[rrnewton]

Finally, with duplicated code between (1) different benchmark modes, and (2) different data structures it's very hard to remember to update each variant and keep everything apples-to-apples. Just be careful.

[rrnewton]

• [x] Minor bug, passing --runs=1 to the Bench1 harness results in:

bench-adaptive-hashmap-1: Prelude.!!: index too large

[vikraman]

adaptive and c-adaptive almost coincide with pure for coldPhase, tested at commit 9dee407. stack bench adaptive-hashmap:bench-adaptive-hashmap-1 '--benchmark-arguments=--doplot --ops=100000000 --bench=cold --runs=5 --minthreads=1 --maxthreads=12 --ratio=5000 --variants=pure --variants=adaptive --variants=c-adaptive'

[rrnewton]

Thanks! Could you post a link to the git revision with the plots? This is a scenario where we keep updating and tweaking so its good to know which version it came from.

[rrnewton]

Since all the checkmarks above are checked off, I was expecting c-adaptive to pull ahead because of some GC savings.. when I can I'll poke at this a bit more to figure out why not. Definitely need to make sure allocation+GCs are happening during that cold phase

@vikraman, when you get a chance I or someone else can show you how to use threadscope if you haven't already used it.