Author: Garret Rieger
Date: March 8th, 2021

Executive Summary

This document investigates the effect of using a variable branch factor and intervals encoded directly in SparseBitSet's on the size of encoded sets. Simulations were ran to determine the reduction or increase in encoded sizes. The results show that a variable branch factor universally reduces sizes with reductions as large as 19%. However, intervals encoded into SparseBitSet's resulted in much smaller reductions, and in some cases increases for large sets.

Based on the results it's recommended to incorporate variable branch factor and intervals into the SparseBitSet specification, but still retain range delta lists as an option for encoding intervals.

Improvements to Sparse Bit Set Encoding

Sparse Bit Set's (described here) are currently used in the patch subset protcol as an efficient means of compressing sets of codepoint values. Currently the sparse bit set's do not efficiently encode large continous ranges of codepoints. So to obtain the smallest encodings they are combined with a separate mechanism for encoding ranges of codepoints.

The existing sparse bit set encoding currently does not ever use zero bytes. This allows them to be used to encode additional information. Since the biggest weakness of the current encoding is inefficient range encoding I came up with a modification to the encoding using the zero byte to efficiently encode ranges:

Any non-leaf node can be encoded as a single zero byte IFF the set contains all values within the interval covered by that node. If a non-leaf node is encoded as a zero byte then all of it's children are emitted from the final encoding.
Since this now allows cases where no leaf nodes get encoded, the final height of the tree can be ambigous. So a single byte is prepended to the front of the encoding that specifies the height of the decoded tree.

An additional weakness of the current sparse bit set encoding is that it uses a fixed branch factor of 8. This was selected because it allowed nodes to be encoded in exactly one byte, however it is unlikely that this is actually the optimal branch factor for the smallest encoding.

Testing the Improvments

To find out if varying the branch factor and/or using zero bytes to encode intervals would produce smaller overall encodings then the existing smallest encoding I reran the set encoding simulations with the updated encodings.

There are four situations that are of interest to us:

encoding a set with no post compression applied (eg. brotli, gzip) and no remapping applied.
encoding a set with no post compression applied, but with codepoint remapping used.
encoding a set with post compression applied, but no remapping applied.
encoding a set with post compression applied and with codepoint remapping used.

The no codepoint remapping situation is of interest because the first request to start a patch subset session will not use codepoint remapping. The no post compression case is of interest because we cannot guarantee which content encoding (if any) the browser will apply to request bodies.

For those 4 situations we want to answer two questions:

Does varying the branch factor used in encoding SparseBitSet's produce a smaller encoding than existing CompressedSets?
Does encoding intervals using a zero byte produce a smaller encoding than existing CompressedSets?

Furthermore we're interested on the effect of the encoding changes across a variety of languages (Chinese, Japanese, and Korean are of highest importance due to the large set sizes needed) and set sizes.

The Results

Finding the Best Branch Factor

To start out I set off to determine what the ideal branch factor would be. I ran the set simulations against sparse bit set encodings for CJK with branch factors varying from 2 to 16. The results were consistent across the languages. Here's an example graph showing the bits needed to encode each codepoint as set size changes for Chinese:

The results showed that a branch factor of 4 produces the smallest encoding for sets less than ~10,000 codepoints beyond that a branch factor of 8 produces the smallest encoding.

I repeated the above simulations this time with zero byte interval encoding enabled (and codepoint remapping enabled since intervals are most useful with remapped codepoints):

The results showed that a branch factor of 4 produces the smallest encoding for sets less than ~5,000 codepoints beyond that a branch factor of 16 produces the smallest encoding.

Based on these results using a branch factor of 4, 8, or 16 produces the smallest encodings depending on the specific situation. For the full set of simulations I used only these 3 branch factors.

Full Simulation Results

The full simulation runs tested:

CompressedSet w/ Branch Factor 8
Minimum of CompressedSet w/ Branch Factor 4, 8, or 16
Minimum of SparseBitSet w/ Intervals and Branch Factor 4, 8, or 16

in the four scenarios of interest:

No codepoint remapping, no brotli post compression.
No codepoint remapping, brotli post compression.
Codepoint remapping, no brotli post compression.
Codepoint remapping, brotli post compression.

against sets of Chinese, Japanese, Korean, and Latin codepoint sets.

Chinese

To see if a variable branch factor could reduce encoding size, the minimum of CompressedSet w/ branch factor 4, 8, or 16 was compared to encoding with only branch factor of 8:

Max Set Size	Reduction in bytes	w/ Brotli	w/ Remapping	w/ Remapping + Brotli
300	19.3%	2.7%	18.0%	2.0%
1,000	17.5%	3.8%	15.3%	5.5%
2,000	15.1%	2.4%	12.1%	3.5%
4,000	12.3%	0.3%	7.5%	2.4%
8,000	7.2%	0.3%	3.7%	0.7%
15,000	1.5%	1.7%	6.4%	0.0%

Varying the branch factor produces reduction in encoded bytes for nearly all set sizes and different encoding scenarios compared to using only branch factor 8.

Next, to see if using intervals in SparseBitSet's is smaller then using CompressedSet's, the minimum of SparseBitSet w/ branch factor 4, 8, or 16 was compared to the minimum of CompressedSet w/ branch factor 4, 8, or 16:

Max Set Size	Reduction in bytes	w/ Brotli	w/ Remapping	w/ Remapping + Brotli
300	0.8%	1.5%	1.0%	1.7%
1,000	0.3%	0.7%	0.4%	0.8%
2,000	0.2%	0.3%	0.2%	0.5%
4,000	0.1%	0.1%	0.2%	0.5%
8,000	0.1%	0.1%	0.1%	2.6%
15,000	0.1%	-0.5%	0.1%	0.7%

Here we see much more modest reductions, typically less then a percent.

Japanese

Minimum of CompressedSet w/ branch factor 4, 8, or 16 versus CompressedSet w/ branch factor 8:

Max Set Size	Reduction in bytes	w/ Brotli	w/ Remapping	w/ Remapping + Brotli
300	17.1%	1.5%	10.0%	8.7%
1,000	14.5%	2.2%	3.9%	3.2%
2,000	12.7%	1.4%	3.6%	3.4%
4,000	9.6%	2.6%	8.6%	1.8%
8,000	4.9%	1.0%	4.8%	2.0%
15,000	0.2%	0.2%	5.3%	0.7%

There's a reduction in bytes across the board with the largest gains for non-brotli compressed sets.

Minimum of SparseBitSet w/ branch factor 4, 8, or 16 versus Minimum of CompressedSet w/ branch factor 4, 8, or 16:

Max Set Size	Reduction in bytes	w/ Brotli	w/ Remapping	w/ Remapping + Brotli
300	1.2%	1.7%	1.8%	1.8%
1,000	0.6%	1.2%	0.7%	1.0%
2,000	-0.2%	0.5%	-1.5%	-1.3%
4,000	-0.1%	-0.5%	-1.8%	-1.0%
8,000	0.2%	0.4%	-1.2%	-1.0%
15,000	-0.3%	0.3%	-4.0%	-3.8%

Here there's very small reductions for small sets (1000 or less codepoints), but for larger sets things get slightly worse.

Korean

Minimum of CompressedSet w/ branch factor 4, 8, or 16 versus CompressedSet w/ branch factor 8:

Max Set Size	Reduction in bytes	w/ Brotli	w/ Remapping	w/ Remapping + Brotli
300	16.8%	2.3%	5.9%	5.7%
1,000	13.5%	0.7%	3.1%	2.5%
2,000	11.7%	0.0%	6.0%	3.7%
4,000	10.2%	0.2%	4.8%	2.6%
8,000	5.9%	0.5%	3.6%	0.6%
15,000	0.8%	1.6%	1.0%	0.9%

There's a reduction in bytes across the board with the largest gains for non-brotli compressed sets.

Minimum of SparseBitSet w/ branch factor 4, 8, or 16 versus Minimum of CompressedSet w/ branch factor 4, 8, or 16:

Max Set Size	Reduction in bytes	w/ Brotli	w/ Remapping	w/ Remapping + Brotli
300	1.2%	1.8%	2.0%	1.9%
1,000	1.6%	1.4%	0.1%	0.2%
2,000	1.5%	1.3%	-0.9%	-1.4%
4,000	0.2%	0.6%	-1.3%	-1.1%
8,000	0.4%	0.4%	-2.1%	-2.1%
15,000	-0.2%	0.4%	-4.2%	-1.6%

For remapped sets there are very small reductions for small sets (1000 or less codepoints), but for larger sets things get slightly worse. For non-remapped sets there's slight reductions for all set sizes (except for the largest).

Latin

Minimum of CompressedSet w/ branch factor 4, 8, or 16 versus CompressedSet w/ branch factor 8:

Max Set Size	Reduction in bytes	w/ Brotli	w/ Remapping	w/ Remapping + Brotli
100	7.9%	6.9%	3.0%	2.5%
300	10.8%	10.1%	1.3%	1.1%

For latin sets there is always reductions in encoding size for all xases when using variable branch factor.

Minimum of SparseBitSet w/ branch factor 4, 8, or 16 versus Minimum of CompressedSet w/ branch factor 4, 8, or 16:

Max Set Size	Reduction in bytes	w/ Brotli	w/ Remapping	w/ Remapping + Brotli
100	8.6%	7.5%	10.8%	9.0%
300	3.3%	3.1%	16.2%	13.1%

Unlike for CJK using SparseBitSet's with intervals for latin produced reductions for all of the scenarios.

Conclusions

Variable Branch Factor

The simulation results show that using a variable branch factor will result in reductions in encoded sizes for nearly all cases. So we should introduce variable branch factor into the specification for SparseBitSet's.

SparseBitSet's with Intervals

For CJK using SparseBitSet's with intervals showed only very small reductions (<1%) and in most cases only for small set sizes. For larger set sizes it caused small regressions. However, for latin there were large reductions across the board.

Given that there are some situations where it can be a useful to reduce encoding sizes versus CompressedSet's it's likely we'll also want to add this to the specification. However, we should also retain the existing delta range list functionality from CompressedSet's as it's still the most efficient means of encoding large intervals.

티스토리 수익 글 보기

티스토리 수익 글 보기

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!