perf: increase min buckets on very small types #524
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
morrisonlevi
force-pushed
the
min-buckets
branch
4 times, most recently
from
192f281 to
f637220
Compare
This was referenced May 8, 2024
Amanieu
reviewed
Jun 17, 2024
|
Also see #47 for previous discussion of this problem. |
Consider `HashSet<u8>` on x86_64 with SSE with various bucket sizes and how many bytes the allocation ends up being: | buckets | capacity | allocated bytes | | ------- | -------- | --------------- | | 4 | 3 | 36 | | 8 | 7 | 40 | | 16 | 14 | 48 | | 32 | 28 | 80 | In general, doubling the number of buckets should roughly double the number of bytes used. However, for small bucket sizes for these small TableLayouts (4 -> 8, 8 -> 16), it doesn't happen. This is an edge case which happens because of padding of the control bytes and adding the Group::WIDTH. Taking the buckets from 4 to 16 (4x) only takes the allocated bytes from 36 to 48 (~1.3x). This platform isn't the only one with edges. Here's aarch64 on an M1 for the same `HashSet<u8>`: | buckets | capacity | allocated bytes | | ------- | -------- | --------------- | | 4 | 3 | 20 | | 8 | 7 | 24 | | 16 | 14 | 40 | Notice 4 -> 8 buckets leading to only 4 more bytes (20 -> 24) instead of roughly doubling. Generalized, `buckets * table_layout.size` needs to be at least as big as `table_layout.ctrl_align`. For the cases I listed above, we'd get these new minimum bucket sizes: - x86_64 with SSE: 16 - aarch64: 8 This is a niche optimization. However, it also removes possible undefined behavior edge case in resize operations. In addition, it may be a useful property to utilize over-sized allocations (see rust-lang#523).
morrisonlevi
force-pushed
the
min-buckets
branch
from
f637220 to
ef4df57
Compare
morrisonlevi
force-pushed
the
min-buckets
branch
from
d00fcaa to
c45e080
Compare
morrisonlevi
commented
Jun 24, 2024
Comment on lines
+216
to
+221
| let cap = cap.max(match (Group::WIDTH, table_layout.size) { | ||
| (16, 0..=1) => 14, | ||
| (16, 2..=3) => 7, | ||
| (8, 0..=1) => 7, | ||
| _ => 3, | ||
| }); |
There was a problem hiding this comment.
It could also be written this way. Not sure what is better:
Suggested change
| let cap = cap.max(match (Group::WIDTH, table_layout.size) { | |
| (16, 0..=1) => 14, | |
| (16, 2..=3) => 7, | |
| (8, 0..=1) => 7, | |
| _ => 3, | |
| }); | |
| let cap = cap.max(match table_layout.size { | |
| 0..=1 if Group::WIDTH == 16 => 14, | |
| 2..=3 if Group::WIDTH == 16 => 7, | |
| 0..=1 if Group::WIDTH == 8 => 7, | |
| _ => 3, | |
| }); |
There was a problem hiding this comment.
What I had in mind was to move the max check inside the if cap < 8 block entirely (and changing the bound to 15). Then just have hard-coded selection depending on the requested capacity and the group width/item size.
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.
Consider
HashSet<u8>on x86_64 with SSE with various bucket sizes and how many bytes the allocation ends up being:In general, doubling the number of buckets should roughly double the number of bytes used. However, for small bucket sizes for these small
TableLayouts (4 -> 8, 8 -> 16), it doesn't happen. This is an edge case which happens because of padding of the control bytes and adding theGroup::WIDTH. Taking the buckets from 4 to 16 (4x) only takes the allocated bytes from 36 to 48 (~1.3x).This platform isn't the only one with edges. Here's aarch64 on an M1 for the same
HashSet<u8>:Notice doubling from 4 to 8 buckets only lead to 4 more bytes (20 -> 24) instead of roughly doubling.
Generalized,
buckets * table_layout.sizeneeds to be at least as big astable_layout.ctrl_align. For the cases I listed above, we'd get these new minimum bucket sizes:This is a niche optimization. However, it also removes possible undefined behavior edge case in resize operations. In addition, it would be a useful property when utilizing over-sized allocations (see #523).