Use insertion sort for small sort_and_dedup_dst_buf inputs
#459
+88
−0
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
I'm still experimenting with the limit, but it's common for sort_and_dedup_dst_buf to be called with a fairly small input array. The bitset approach's cost is proportional to the max - min value delta, so when there's only a few entries it may be faster to do insertion sort instead -- 16 entries will probably fit in a single cache line. Changing SMALL_INPUT_LIMIT changes cutoff, commenting it out disables that approach outright. Setting SMALL_INPUT_CHECK to 1 enables a pass at the end that verifies the array is actually sorted and unique. Also, check upfront whether the input array is already sorted and unique. Experiments so far suggest that this doesn't pay off very often, but it's very cheap since we're already sweeping over the input to the min and max values.
I benchmarked with a range of values 8 to 4096, somewhere around 32 - 256 worked particularly well, and it started to become slower beyond that, so set it to <32 for now.
|
As of this PR: |
katef
deleted the
sv/use-insertion-sort-for-small-sort-and-dedup-inputs
branch
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
As a follow up to #456 -- kate mentioned that
examples/wordsstill had the same outliers (#456 (comment)). As hashing dropped lower in the profile, spending a disproportionate amount of time insort_and_dedup_dst_bufshowed up. This makes sense, because the outliers in thewordsoutput come from blab's successive seeds generating specific outputs (such as combining 855 instances of the regexGyADewtfILV53SSPua4IKbEZbug6BWTJ8o22K6ydeXs0NWsMADsGyADewtfILV53SSPua4IKbEZbug6BWTJ8o22K6ydeXs0NWsMA...) that are pathological knots for determinisation to untangle, not just a gradually increasing input size.Currently,
sort_and_dedup_dst_bufhandles its input with a couple strategies:The outlier appeared to be due to from frequently calling it with inputs which only had a few values, but a large delta between the min and max values, so it was effectively zeroing a couple KB on the stack, setting just a few bits, and then sweeping over it again to write out an array of the set bits -- when the input array is fairly small (
< = SMALL_INPUT_LIMIT), it's faster to append ascending values into a temporary buffer (so already sorted input just copies over) and then shift it forward as necessary to plug in the remaining values. In other words, use insertion sort when the entire data set fits within a few cache lines. Iit often does.When using a
words.shoutput test data set kate sent me for benchmarking, this brings the total runtime forwords -dtfrom about 11.7 seconds to about 8.4 seconds on my laptop.I also added a check for whether the input is already sorted and deduplicated, so the function doesn't need to do anything. During benchmarking that case was only reached a small percent of the time, but it's very cheap to also note that while sweeping through to find the min and max values.