Travis
June 24, 2026, 2:34am
21
That seems to be a good idea. Iβve done this along with adding automatic compaction when blocks are shrunk and now were around 90% of croaringβs speed on this benchmark
Of course that means 2 allocations instead of one and a slightly less happy deinit . But I can accept that.
zig build bench -Doptimize=ReleaseFast
$ zig build bench -Doptimize=ReleaseFast
--------------------------------------------------
overall:
--------------------------------------------------
CRoaring: 11430 ops 20.698ms 0.55MB ops/sec
ZRoaring: 11430 ops 22.894ms 0.50MB ops/sec
ratio -- 0.90 π₯
--------------------------------------------------
individual ops: speed=ops/sec
--------------------------------------------------
op cr speed zr speed # ratio
--------------------------------------------------
clear 1.62MB 35.25MB 240 21.82 β‘
run_optimize 0.61MB 0.60MB 365 0.99 ππ»
shrink_to_fit 4.17MB 1.20MB 330 0.29 π©
portable_serialize 0.70MB 0.94MB 105 1.34 β‘
frozen_serialize 2.31MB 2.28MB 85 0.99 ππ»
minimum 32.69MB 34.56MB 580 1.06 ππ»
maximum 14.55MB 7.65MB 585 0.53 π©
add 12.09MB 10.37MB 2200 0.86 π₯
rank 10.29MB 9.69MB 25 0.94 ππ»
select 8.72MB 8.43MB 30 0.97 ππ»
contains 22.46MB 22.67MB 580 1.01 ππ»
add_many 2.52MB 2.86MB 915 1.13 β‘
add_range_closed 2.37MB 2.29MB 1545 0.96 ππ»
contains_range 16.02MB 17.40MB 570 1.09 β‘
range_cardinality 6.67MB 5.15MB 5 0.77 π₯
remove 11.65MB 11.98MB 630 1.03 ππ»
and 0.34MB 0.41MB 420 1.22 β‘
or 196.69kB 219.31kB 305 1.12 β‘
xor 157.98kB 180.81kB 375 1.14 β‘
andnot 217.86kB 0.33MB 320 1.50 β‘
lazy_or 161.72kB 239.00kB 250 1.48 β‘
or_inplace 0.32MB 0.34MB 390 1.05 ππ»
and_inplace 140.05kB 142.59kB 60 1.02 ππ»
is_subset 134.90kB 144.89kB 180 1.07 β‘
equals 26.55MB 6.86MB 30 0.26 π©
and_cardinality 0.74MB 1.05MB 40 1.42 β‘
or_cardinality 0.79MB 0.93MB 45 1.18 β‘
xor_cardinality 0.95MB 0.94MB 35 0.99 ππ»
andnot_cardinality 0.81MB 0.96MB 20 1.18 β‘
jaccard_index 0.88MB 1.06MB 40 1.20 β‘
or_many 19.59kB 12.23kB 130 0.62 π©
1 Like
Travis
June 29, 2026, 12:36am
22
Iβve just pushed some minor optimizations which bring the overall benchmark iβve been sharing up around 1% to 91% from 90% last time.
More importantly, Iβve improved the overall benchmark a bit. It now uses clear (which is an alias to deinit) instead of clear_retaining_capacity.
As a result clear is no longer like 20x faster
Overall plot:
$ zig build bench -Doptimize=ReleaseFast
$ zig build bench -Doptimize=ReleaseFast
----------------------------------------------------
overall:
----------------------------------------------------
CRoaring: 23450 ops 46.171ms 0.508MB ops/sec
ZRoaring: 23450 ops 50.605ms 0.463MB ops/sec
ratio -- 0.912 π₯
----------------------------------------------------
individual ops: speed=ops/sec
----------------------------------------------------
op cr speed zr speed # ratio
----------------------------------------------------
clear 1.62MB 1.66MB 530 1.03 ππ»
run_optimize 0.46MB 0.40MB 780 0.88 π₯
shrink_to_fit 4.21MB 3.08MB 700 0.73 π©
portable_serialize 3.07MB 2.32MB 210 0.75 π₯
portable_deserialize 216.75kB 0.42MB 860 1.94 β‘
frozen_serialize 3.03MB 2.39MB 170 0.79 π₯
minimum 27.32MB 26.72MB 970 0.98 ππ»
maximum 21.53MB 18.53MB 950 0.86 π₯
add 12.05MB 12.03MB 4470 1.00 ππ»
rank 9.69MB 9.69MB 50 1.00 ππ»
select 8.72MB 4.50MB 60 0.52 π©
contains 23.15MB 22.07MB 800 0.95 ππ»
add_many 2.52MB 3.11MB 1900 1.24 β‘
add_range_closed 2.46MB 2.21MB 3210 0.90 π₯
contains_range 19.16MB 18.89MB 1050 0.99 ππ»
range_cardinality 8.33MB 5.52MB 10 0.66 π©
remove 12.06MB 12.43MB 1360 1.03 ππ»
and 0.26MB 0.27MB 930 1.04 ππ»
or 201.05kB 189.69kB 610 0.94 ππ»
xor 155.58kB 183.67kB 750 1.18 β‘
andnot 212.07kB 0.43MB 640 2.04 β‘
lazy_or 168.66kB 226.22kB 500 1.34 β‘
or_inplace 0.31MB 0.33MB 780 1.08 β‘
and_inplace 147.68kB 154.56kB 120 1.05 ππ»
is_subset 134.64kB 139.47kB 360 1.04 ππ»
equals 28.99MB 20.62MB 60 0.71 π©
and_cardinality 0.66MB 0.77MB 80 1.16 β‘
or_cardinality 0.75MB 0.76MB 90 1.03 ππ»
xor_cardinality 0.93MB 0.91MB 70 0.98 ππ»
andnot_cardinality 0.70MB 0.82MB 40 1.17 β‘
jaccard_index 0.80MB 0.79MB 80 0.98 ππ»
or_many 20.50kB 11.80kB 260 0.58 π©
1 Like
Travis
July 7, 2026, 1:31am
23
Excited to see zroaring has pulled even with CRoaring on the overall ReleaseFast benchmark iβve been sharing.
This is the result of 2 recent changes:
dropping the resizable-struct dependency for a normal extern struct with faster, static offset calculations.
moving Block storage from Bitmap to Container like CRoaring does. while this complicates memory management it makes for fewer required copies and scales better as Bitmaps grow large.
Iβd still like to find a a solution with less pointers chasing in deinit, maybe storing Blocks together somehow. But for now Iβm happy with the perf improvements.
$ zig build bench -Doptimize=ReleaseFast
----------------------------------------------------
overall:
----------------------------------------------------
CRoaring: 29210 ops 147.571ms 197.937kB ops/sec
ZRoaring: 29210 ops 146.932ms 198.798kB ops/sec
ratio -- 1.004 ππ»
----------------------------------------------------
individual ops: speed=ops/sec
----------------------------------------------------
op cr speed zr speed # ratio
----------------------------------------------------
clear 1.02MB 0.79MB 830 0.78 π₯
run_optimize 229.32kB 179.44kB 1170 0.78 π₯
shrink_to_fit 3.58MB 4.61MB 980 1.29 β‘
portable_serialize 0.49MB 0.40MB 530 0.82 π₯
frozen_serialize 0.57MB 0.49MB 380 0.87 π₯
minimum 18.34MB 18.73MB 280 1.02 ππ»
maximum 15.22MB 11.95MB 210 0.79 π₯
add 10.72MB 8.39MB 4780 0.78 π₯
rank 8.03MB 7.09MB 410 0.88 π₯
select 4.36MB 3.34MB 400 0.77 π₯
contains 13.53MB 14.10MB 300 1.04 ππ»
add_many 2.44MB 2.08MB 2290 0.85 π₯
add_range_closed 2.52MB 1.86MB 3540 0.74 π©
contains_range 9.65MB 8.80MB 330 0.91 π₯
range_cardinality 11.89MB 12.43MB 390 1.05 ππ»
remove 11.95MB 12.52MB 1610 1.05 ππ»
and 224.95kB 232.43kB 1210 1.03 ππ»
or 109.21kB 117.00kB 880 1.07 β‘
xor 86.01kB 102.43kB 1070 1.19 β‘
andnot 103.16kB 132.21kB 950 1.28 β‘
lazy_or 98.76kB 91.53kB 840 0.93 π₯
or_inplace 200.11kB 207.22kB 1160 1.04 ππ»
and_inplace 125.61kB 100.96kB 320 0.80 π₯
is_subset 28.69kB 30.32kB 710 1.06 ππ»
equals 2.94MB 7.80MB 320 2.65 β‘
and_cardinality 0.45MB 0.42MB 420 0.94 ππ»
or_cardinality 0.45MB 0.42MB 400 0.93 ππ»
xor_cardinality 0.74MB 0.68MB 280 0.91 π₯
andnot_cardinality 0.36MB 0.33MB 240 0.92 π₯
jaccard_index 0.72MB 0.63MB 290 0.88 π₯
or_many 13.47kB 12.67kB 610 0.94 ππ»
portable_deserialize 44.32kB 67.20kB 330 1.52 β‘
statistics 2.43MB 2.08MB 300 0.86 π₯
flip 163.00kB 118.95kB 450 0.73 π©
Vind wasm binary size went down a bit too:
$ ~/.../zig/vind]$ zig build wasm -Droaring=zig && ls -lh zig-out/wasm/vind.wasm
... 113K Jul 6 17:29
Vind bench seems around the same here:
$ zig build bench -Droaring=c
timings
build 686.287 ms 72856 docs/s
finalize 7.506 ms 6660895 docs/s
serialize 5.171 ms 1379200769 bytes/s
deserialize 700.814 ms 71346 docs/s
total 1399.778 ms
...
combined total 2165.707 ms
$ zig build bench -Droaring=zig
timings
build 699.478 ms 71482 docs/s
finalize 7.497 ms 6668907 docs/s
serialize 5.262 ms 1355371566 bytes/s
deserialize 717.759 ms 69661 docs/s
total 1429.997 ms
...
combined total 2184.581 ms
6 Likes
Excellent numbers! Thank you for the update!
2 Likes
Travis
July 11, 2026, 5:18am
25
More changes to reduce the number of allocations.
Previously Array.containers field was a [*]*Container. No good reason for the extra indirection except that I was following CRoaring. And now its [*]Container.
Previously Container.blocks was a separate allocation. And now blocks are allocated together with each container.
So Containers are now allocated (like Array) with a flexible blocks field. I noticed my allocation failures test time is about half what it was before. This change has improved most measurements in the following benchmark, most notably cache misses.
$ poop zig-out/bin/bench-cr zig-out/bin/bench-zr
zig build -Dbench-target=cr -Doptimize=ReleaseFast && zig build -Dbench-target=zr -Doptimize=ReleaseFast && poop zig-out/bin/bench-cr zig-out/bin/bench-zr
Benchmark 1 (575 runs): zig-out/bin/bench-cr
measurement mean Β± Ο min β¦ max outliers delta
wall_time 8.64ms Β± 278us 8.09ms β¦ 11.5ms 9 ( 2%) 0%
peak_rss 5.94MB Β± 100KB 5.64MB β¦ 6.26MB 72 (13%) 0%
cpu_cycles 27.0M Β± 446K 26.5M β¦ 33.4M 25 ( 4%) 0%
instructions 69.3M Β± 3.46 69.3M β¦ 69.3M 28 ( 5%) 0%
cache_references 2.24M Β± 22.9K 2.18M β¦ 2.36M 21 ( 4%) 0%
cache_misses 221K Β± 4.52K 206K β¦ 241K 13 ( 2%) 0%
branch_misses 130K Β± 1.94K 126K β¦ 137K 1 ( 0%) 0%
Benchmark 2 (603 runs): zig-out/bin/bench-zr
measurement mean Β± Ο min β¦ max outliers delta
wall_time 8.25ms Β± 199us 7.79ms β¦ 9.11ms 7 ( 1%) β‘- 4.5% Β± 0.3%
peak_rss 5.47MB Β± 64.1KB 5.19MB β¦ 5.61MB 19 ( 3%) β‘- 8.0% Β± 0.2%
cpu_cycles 26.8M Β± 323K 26.4M β¦ 29.9M 27 ( 4%) - 0.7% Β± 0.2%
instructions 69.4M Β± 3.70 69.4M β¦ 69.4M 15 ( 2%) + 0.2% Β± 0.0%
cache_references 2.34M Β± 29.1K 2.26M β¦ 2.57M 22 ( 4%) π©+ 4.5% Β± 0.1%
cache_misses 216K Β± 4.27K 205K β¦ 239K 9 ( 1%) β‘- 2.3% Β± 0.2%
branch_misses 110K Β± 1.34K 107K β¦ 115K 15 ( 2%) β‘- 15.5% Β± 0.1%
This is the first time weβve been faster than CRoaring
$ zig build bench -Doptimize=ReleaseFast
$ zig build bench -Doptimize=ReleaseFast -- write_csv_row
----------------------------------------------------
overall:
----------------------------------------------------
CRoaring: 30520 ops 97.042ms 0.315MB ops/sec
ZRoaring: 30520 ops 96.418ms 0.317MB ops/sec
ratio -- 1.006 ππ»
----------------------------------------------------
individual ops: speed=ops/sec
----------------------------------------------------
op cr speed zr speed # ratio
----------------------------------------------------
clear 0.67MB 0.96MB 990 1.42 β‘
run_optimize 0.42MB 0.34MB 1200 0.80 π₯
shrink_to_fit 4.30MB 5.23MB 940 1.22 β‘
portable_serialize 1.54MB 1.09MB 540 0.71 π©
frozen_serialize 3.50MB 2.39MB 420 0.68 π©
minimum 23.49MB 26.29MB 230 1.12 β‘
maximum 17.68MB 16.55MB 290 0.94 ππ»
add 11.58MB 11.71MB 5150 1.01 ππ»
rank 11.30MB 11.01MB 390 0.97 ππ»
select 8.28MB 6.97MB 360 0.84 π₯
contains 17.33MB 16.74MB 200 0.97 ππ»
add_many 2.65MB 2.54MB 2280 0.96 ππ»
add_range_closed 2.64MB 2.19MB 3690 0.83 π₯
contains_range 15.34MB 15.91MB 360 1.04 ππ»
range_cardinality 14.84MB 15.85MB 330 1.07 ππ»
remove 13.86MB 14.44MB 1760 1.04 ππ»
and 0.30MB 0.31MB 1190 1.03 ππ»
or 191.93kB 196.96kB 950 1.03 ππ»
xor 167.31kB 162.64kB 1170 0.97 ππ»
andnot 194.80kB 0.28MB 1010 1.44 β‘
lazy_or 137.32kB 117.23kB 920 0.85 π₯
or_inplace 0.31MB 0.27MB 1090 0.85 π₯
and_inplace 0.33MB 0.33MB 540 1.01 ππ»
is_subset 182.30kB 197.39kB 650 1.08 β‘
equals 9.81MB 25.78MB 330 2.63 β‘
and_cardinality 2.37MB 2.26MB 300 0.95 ππ»
or_cardinality 1.42MB 1.47MB 380 1.03 ππ»
xor_cardinality 1.46MB 1.45MB 350 0.99 ππ»
andnot_cardinality 1.33MB 1.34MB 240 1.01 ππ»
jaccard_index 0.48MB 0.53MB 280 1.11 β‘
or_many 12.50kB 13.44kB 570 1.08 β‘
portable_deserialize 154.42kB 195.40kB 340 1.27 β‘
statistics 4.89MB 5.10MB 310 1.04 ππ»
flip 189.46kB 93.35kB 440 0.49 π©
frozen_view 1.63MB 1.00MB 330 0.61 π©
@xsawyerx weβre sligntly faster than CRoaring now on your benchmark
Vind benchmark
[~/.../zig/vind]$ zig build bench -Droaring=c
timings
build 753.437 ms 66363 docs/s
finalize 7.962 ms 6280130 docs/s
serialize 5.351 ms 1332837062 bytes/s
deserialize 740.708 ms 67503 docs/s
total 1507.458 ms
...
combined total 2299.986 ms
[~/.../zig/vind]$ zig build bench -Droaring=zig
timings
build 712.943 ms 70132 docs/s
finalize 7.503 ms 6664204 docs/s
serialize 5.288 ms 1348710113 bytes/s
deserialize 724.253 ms 69037 docs/s
total 1449.987 ms
...
combined total 2214.106 ms
4 Likes
Very impressive! I guess Vind is moving to ZRoaring!
1 Like
.
. And the overall benchmark is now a more honest comparison with CRoaring.
. I havenβt been sharing this one here, but previously CRoaring was between 5% and 15% faster on this one in the past month.
. Now you have no reason to use CRoaring except for binary size 
.
