How much linear memory access is enough?
Comments
sweetjuly
gwking
I’ve casually experimented with this in python a number of times for various hot loops, including those where I’m passing the chunk between c routines. On Apple M1 I’ve never seen a case where chunks larger than 16k mattered. That’s the page size, so totally unsurprising.
Nevertheless it’s been a helpful rule of thumb to not overthink optimizations.
PhilipTrettner
I looked into this because part of our pipeline is forced to be chunked. Most advice I've seen boils down to "more contiguity = better", but without numbers, or at least not generalizable ones.
My concrete tasks will already reach peak performance before 128 kB and I couldn't find pure processing workloads that benefit significantly beyond 1 MB chunk size. Code is linked in the post, it would be nice to see results on more systems.
twoodfin
Your results match similar analyses of database systems I’ve seen.
64KB-128KB seems like the sweet spot.
smj-edison
Side note, but this product looks really cool! I have a fundamental mistrust of all boolean operations, so to see a system that actually works with degenerate cases correctly is refreshing.
aapoalas
Would kernel huge pages possibly have an effect here also?
_zoltan_
is this an attempt at nerd sniping? ;-)
on GPU databases sometimes we go up to the GB range per "item of work" (input permitting) as it's very efficient.
I need to add it to my TODO list to have a look at your github code...
PhilipTrettner
It definitely worked on myself :)
Do have a look, I've tried to roughly keep it small and readable. It's ~250 LOC effectively.
Also, this is CPU only. I'm not super sure what a good GPU version of my benchmark would be, though ... Maybe measuring a "map" more than a "reduction" like I do on the CPU? We should probably take a look at common chunking patterns there.
01HNNWZ0MV43FF
This is good data, but I'm not sure what the actionable is for me as a Grug Programmer.
It means if I'm doing very light processing (sums) I should try to move that to structure-of-arrays to take advantage of cache? But if I'm doing something very expensive, I can leave it as array-of-structures, since the computation will dominate the memory access in Amdahl's Law analysis?
This data should tell me something about organizing my data and accessing it, right?
I wonder how much of the cost is coming from the cache misses vs the more frequent indirections/ILP drop?
For example, I wonder what this test looks like if you don't randomize the chunks but instead just have the chunks in work order? If you still see the perf hit, that suggests the cost is not from the cache misses but rather the overhead of needing to switch chunks more often.