jacobi

JACOBI Project Benchmark Report (Feb 2026)

Build Instructions
Benchmark Dataset
Benchmark Settings
Benchmark Raw Results
Benchmark Processed Results
Results Analysis
- Rank of Order Book implementations
- Visual Reports
Reflection on the Results

JACOBI explores various design approaches for implementing high-performance, low-latency Order Books (L3) in C++.

This document details a comprehensive case study running the JACOBI benchmark suite against actual market data (streams of adds, modifies, and executes) to compare implementations strategies.

The experiment was executed following the Running a Benchmark Experiment guidelines using tag benchmark-feb2026.

Hardware and Environment

System details:

$ uname -a
Linux jacobi.benchmark2 6.8.0-100-generic #100-Ubuntu SMP PREEMPT_DYNAMIC Tue Jan 13 16:40:06 UTC 2026 x86_64 x86_64 x86_64 GNU/Linux

$ cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
performance

$ cat /proc/cpuinfo
//.. core 2 is used for running benchmark apps.
processor       : 2
vendor_id       : AuthenticAMD
cpu family      : 25
model           : 97
model name      : AMD Ryzen 9 7950X3D 16-Core Processor
stepping        : 2
microcode       : 0xa601209
cpu MHz         : 4989.492
cache size      : 1024 KB

$ cat /sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_max_freq
5759000

$ lscpu | egrep 'Model name|L1d cache|L1i cache|L2 cache|L3 cache'
Model name:                              AMD Ryzen 9 7950X3D 16-Core Processor
L1d cache:                               512 KiB (16 instances)
L1i cache:                               512 KiB (16 instances)
L2 cache:                                16 MiB (16 instances)
L3 cache:                                128 MiB (2 instances)

# 512 KiB (16 instances) means 512/16=32 KiB per Core
# 16 MiB (16 instances) means 16/16=1 MiB per Core

$ cat /proc/meminfo | fgrep MemTotal
MemTotal:       97949508 kB     # 96GB

The benchmark suite was compiled using two industry-standard compilers:

$ g++-14 --version
g++-14 (Ubuntu 14.2.0-4ubuntu2~24.04) 14.2.0

$ clang++-20 --version
Ubuntu clang version 20.1.2 (0ubuntu1~24.04.2)

# Using `-stdlib=libc++` for clang.

Build Instructions

Both builds were executed in Release mode using Conan for dependency management. Crucially, the -march=znver4 flag was passed via the Conan profile to enable Zen 4 specific instruction sets:

# gcc build:
conan install -pr:a ./conan_profiles/ubu-gcc-14-znver4 -s:a build_type=Release --build missing -of _gcc14_release .
( source ./_gcc14_release/conanbuild.sh && cmake -B_gcc14_release . -DCMAKE_TOOLCHAIN_FILE=_gcc14_release/conan_toolchain.cmake -DCMAKE_BUILD_TYPE=Release)
cmake --build _gcc14_release

# clang build:
conan install -pr:a ./conan_profiles/ubu-clang-20-znver4 -s:a build_type=Release --build missing -of _clang20_release .
( source ./_clang20_release/conanbuild.sh && cmake -B_clang20_release . -DCMAKE_TOOLCHAIN_FILE=_clang20_release/conan_toolchain.cmake -DCMAKE_BUILD_TYPE=Release)
cmake --build _clang20_release

# It also makes sense to run tests:
ctest -T test --timeout 60 --output-on-failure --test-dir _gcc14_release
ctest -T test --timeout 60 --output-on-failure --test-dir _clang20_release

Benchmark Dataset

Binary input files containing dense streams of book events were prepared from historical market data. To ensure comprehensive coverage, the datasets are split into two categories:

_snapshots contains single book events files;
_snapshots_multi contains events for multiple books.

Benchmark Settings

The following settings files were used:

Benchmark Raw Results

Note: The original results were post-processed to hide filenames containing private tags.

Benchmark Processed Results

Benchmark results were converted to csv and sqlite db files.

CSV:

python ./benchmark_experiment/results2db.py --commit 437dacf --compiler gcc14 --json_path ./docs/feb2026/raw_results/benchmark_gcc14.single.range.json --output ./docs/feb2026/results/benchmark_gcc14.single.range.csv
python ./benchmark_experiment/results2db.py --commit 437dacf --compiler gcc14 --json_path ./docs/feb2026/raw_results/benchmark_gcc14.multi.range.json --output ./docs/feb2026/results/benchmark_gcc14.multi.range.csv
python ./benchmark_experiment/results2db.py --commit 437dacf --compiler clang20 --json_path ./docs/feb2026/raw_results/benchmark_clang20.single.range.json --output ./docs/feb2026/results/benchmark_clang20.single.range.csv
python ./benchmark_experiment/results2db.py --commit 437dacf --compiler clang20 --json_path ./docs/feb2026/raw_results/benchmark_clang20.multi.range.json --output ./docs/feb2026/results/benchmark_clang20.multi.range.csv

All csv tables

SQLite:

# Separate tables:
python ./benchmark_experiment/results2db.py --commit 437dacf --compiler gcc14 --json_path ./docs/feb2026/raw_results/benchmark_gcc14.single.range.json --table gcc14_single --output ./docs/feb2026/results/benchmarks_all.db
python ./benchmark_experiment/results2db.py --commit 437dacf --compiler gcc14 --json_path ./docs/feb2026/raw_results/benchmark_gcc14.multi.range.json --table gcc14_multi --output ./docs/feb2026/results/benchmarks_all.db
python ./benchmark_experiment/results2db.py --commit 437dacf --compiler clang20 --json_path ./docs/feb2026/raw_results/benchmark_clang20.single.range.json --table clang20_single --output ./docs/feb2026/results/benchmarks_all.db
python ./benchmark_experiment/results2db.py --commit 437dacf --compiler clang20 --json_path ./docs/feb2026/raw_results/benchmark_clang20.multi.range.json --table clang20_multi --output ./docs/feb2026/results/benchmarks_all.db

# Merged results for both compilers (multi/single are still separate):
python ./benchmark_experiment/results2db.py --commit 437dacf --compiler gcc14 --json_path ./docs/feb2026/raw_results/benchmark_gcc14.single.range.json --table all_single --output ./docs/feb2026/results/benchmarks_all.db
python ./benchmark_experiment/results2db.py --commit 437dacf --compiler gcc14 --json_path ./docs/feb2026/raw_results/benchmark_gcc14.multi.range.json --table all_multi --output ./docs/feb2026/results/benchmarks_all.db
python ./benchmark_experiment/results2db.py --commit 437dacf --compiler clang20 --json_path ./docs/feb2026/raw_results/benchmark_clang20.single.range.json --table all_single --output ./docs/feb2026/results/benchmarks_all.db
python ./benchmark_experiment/results2db.py --commit 437dacf --compiler clang20 --json_path ./docs/feb2026/raw_results/benchmark_clang20.multi.range.json --table all_multi --output ./docs/feb2026/results/benchmarks_all.db

SQLite db

Contains the following tables:

gcc14_single Measurements for gcc-14 single book benchmarks.
gcc14_multi Measurements for gcc-14 multi book benchmarks.
clang20_single Measurements for clang-20 single book benchmarks.
clang20_multi Measurements for clang-20 multi book benchmarks.
all_single Measurements for all compilers single book benchmarks.
all_multi Measurements for all compilers multi book benchmarks.

Results Analysis

To build a rank of implementations we use the following query:

-- Use one of the table-names defined above:
WITH measurements AS ( SELECT * FROM gcc14_single ),
-- Single books data set pXX params:
     pxx_params AS (SELECT 2 AS p99_offset, 29 AS p90_offset, 150 AS p50_offset)
-- Multi books data set pXX params:
     -- pxx_params AS (SELECT 0 AS p99_offset, 3 AS p90_offset, 16 AS p50_offset)

SELECT
    -- ==========================================
    -- Implementation:
    compiler,
    orders_table_storage,
    components,
    -- ==========================================

    avg(mps) as avg_mps,
    min(mps) as min_mps,
    max(mps) as max_mps,

    -- ==========================================
    -- PXX results:
    (SELECT mps
        FROM measurements as mea2
            WHERE mea2.orders_table_storage = main.orders_table_storage
                  and mea2.components = main.components
                  and mea2.compiler = main.compiler
                ORDER BY mea2.mps
        LIMIT 1 OFFSET (SELECT p99_offset FROM pxx_params ) )
    as p99,

    (SELECT mps
        FROM measurements as mea2
            WHERE mea2.orders_table_storage = main.orders_table_storage
                  and mea2.components = main.components
                  and mea2.compiler = main.compiler
                ORDER BY mea2.mps
        LIMIT 1 OFFSET (SELECT p90_offset FROM pxx_params ) )
    as p90,

    (SELECT mps
        FROM measurements as mea2
            WHERE mea2.orders_table_storage = main.orders_table_storage
                  and mea2.components = main.components
                  and mea2.compiler = main.compiler
                ORDER BY mea2.mps
        LIMIT 1 OFFSET (SELECT p50_offset FROM pxx_params ) )
    as p50,
    -- ==========================================

    -- ==========================================
    -- Which files yield worst results for this implementation,
    -- use it for further research:
    (SELECT jacobi_data
        FROM measurements as mea2
            WHERE mea2.orders_table_storage = main.orders_table_storage
                  and mea2.components = main.components
                  and mea2.compiler = main.compiler
                ORDER BY mea2.mps
        LIMIT 1 OFFSET 0 ) as worst_0,

    (SELECT jacobi_data
        FROM measurements as mea2
            WHERE mea2.orders_table_storage = main.orders_table_storage
                  and mea2.components = main.components
                  and mea2.compiler = main.compiler
                ORDER BY mea2.mps
        LIMIT 1 OFFSET 1 ) as worst_1,

    (SELECT jacobi_data
        FROM measurements as mea2
            WHERE mea2.orders_table_storage = main.orders_table_storage
                  and mea2.components = main.components
                  and mea2.compiler = main.compiler
                ORDER BY mea2.mps
        LIMIT 1 OFFSET 2 ) as worst_2
    -- ==========================================

FROM measurements as main
GROUP BY orders_table_storage, components, compiler
ORDER BY
    -- Weighted importance of characteristics:
    avg_mps + p99 + p90 + p50
    -- Or alternatively:
    -- avg_mps
    -- p99
    -- p90
    -- p50
desc

Rank of Order Book implementations

Single book handling gcc-14 implementations (TOP 20):

orders_table_storage	components	avg_mps	min_mps	max_mps	p99	p90	p50	worst_0	worst_1	worst_2
mixed_hot_cold_h1024	`bsn1_plvl30_refIX3`	44.2884	1.88874	55.1051	9.5846	37.9558	45.5046	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
mixed_hot_cold_h256	`bsn1_plvl30_refIX3`	43.8449	2.11236	55.3187	3.03278	37.5838	45.2059	284_238k.jacobi_data	259_300k.jacobi_data	285_166k.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl31_refIX3`	43.8	1.84927	54.1279	9.38197	37.5309	44.8927	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
linear_v2	`bsn1_plvl31_refIX3`	43.7199	0.301706	59.5851	1.91619	38.0224	45.5224	273_148k.jacobi_data	083_694k.jacobi_data	241_100k.jacobi_data
mixed_hot_cold_h256	`bsn1_plvl31_refIX3`	43.3018	2.04677	55.5127	2.88552	37.3653	44.5804	284_238k.jacobi_data	259_300k.jacobi_data	285_166k.jacobi_data
linear_v2	`bsn1_plvl30_refIX3`	43.1368	0.885178	56.1959	4.67282	36.774	44.554	273_148k.jacobi_data	083_694k.jacobi_data	241_100k.jacobi_data
mixed_hot_cold_h128	`bsn1_plvl30_refIX3`	42.7437	2.78716	55.0785	4.35123	35.3639	44.3506	284_238k.jacobi_data	259_300k.jacobi_data	285_166k.jacobi_data
mixed_hot_cold_h128	`bsn1_plvl31_refIX3`	42.3088	2.6317	54.94	4.27141	35.6389	43.7411	284_238k.jacobi_data	259_300k.jacobi_data	285_166k.jacobi_data
mixed_hot_cold_h64	`bsn1_plvl30_refIX3`	41.4206	4.40493	55.6178	6.63583	32.8821	43.0492	284_238k.jacobi_data	259_300k.jacobi_data	285_166k.jacobi_data
linear_v1	`bsn1_plvl31_refIX3`	41.377	0.304808	53.4187	1.73483	33.8081	43.9138	273_148k.jacobi_data	083_694k.jacobi_data	186_1944k.jacobi_data
mixed_hot_cold_h64	`bsn1_plvl31_refIX3`	41.3375	4.24288	54.1262	6.62813	33.4906	42.7965	284_238k.jacobi_data	259_300k.jacobi_data	285_166k.jacobi_data
linear_v2	`bsn1_plvl51_refIX3`	41.0562	2.34565	50.4295	17.2688	31.8796	42.3675	273_148k.jacobi_data	083_694k.jacobi_data	186_1944k.jacobi_data
linear_v2	`bsn1_plvl12_refIX3`	40.6218	0.875325	55.4466	5.03132	34.5131	41.1963	273_148k.jacobi_data	083_694k.jacobi_data	241_100k.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl12_refIX3`	40.5528	1.83727	52.8096	9.55078	35.376	41.168	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
linear_v2	`bsn1_plvl41_refIX3`	40.3976	1.16372	53.2686	9.30627	27.692	42.5029	273_148k.jacobi_data	083_694k.jacobi_data	241_100k.jacobi_data
linear_v1	`bsn1_plvl51_refIX3`	40.3066	2.35328	50.2622	10.3583	30.0912	42.1778	273_148k.jacobi_data	083_694k.jacobi_data	186_1944k.jacobi_data
mixed_hot_cold_h256	`bsn1_plvl12_refIX3`	40.0871	2.06089	53.3618	2.97222	34.9382	40.7926	284_238k.jacobi_data	259_300k.jacobi_data	285_166k.jacobi_data
linear_v2	`bsn1_plvl11_refIX3`	39.9833	2.4581	51.4612	14.9472	35.6174	40.4675	273_148k.jacobi_data	083_694k.jacobi_data	241_100k.jacobi_data
linear_v2	`bsn1_plvl42_refIX3`	39.9209	0.611602	50.8122	7.74623	29.4512	41.537	273_148k.jacobi_data	083_694k.jacobi_data	065_125k.jacobi_data
mixed_hot_cold_h32	`bsn1_plvl31_refIX3`	39.9022	6.92707	51.0068	10.2256	31.2759	41.2008	284_238k.jacobi_data	259_300k.jacobi_data	285_166k.jacobi_data

Single book handling clang-20 implementations (TOP 20):

orders_table_storage	components	avg_mps	min_mps	max_mps	p99	p90	p50	worst_0	worst_1	worst_2
linear_v2	`bsn1_plvl11_refIX3`	44.3905	3.08987	58.457	22.3295	38.4883	44.9381	273_148k.jacobi_data	083_694k.jacobi_data	186_1944k.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl11_refIX3`	43.9256	5.56357	54.0438	21.6776	38.3558	44.6825	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
linear_v2	`bsn1_plvl51_refIX3`	45.1507	2.55143	54.4846	19.4928	36.7451	46.3802	273_148k.jacobi_data	083_694k.jacobi_data	186_1944k.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl51_refIX3`	44.8569	5.05421	54.9282	18.6164	36.1017	46.0817	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl30_refIX3`	46.8915	1.9287	58.5714	10.617	39.6212	47.8927	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl31_refIX3`	46.1905	1.66626	55.1726	8.9965	41.2374	47.4413	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
linear_v2	`bsn1_plvl30_refIX3`	47.6105	0.883425	58.2893	4.83673	41.1537	49.1682	273_148k.jacobi_data	083_694k.jacobi_data	241_100k.jacobi_data
linear_v1	`bsn1_plvl11_refIX3`	43.9377	3.12486	56.701	13.8634	38.2099	44.6046	273_148k.jacobi_data	083_694k.jacobi_data	186_1944k.jacobi_data
linear_v1	`bsn1_plvl51_refIX3`	45.3074	2.55738	55.4899	11.6867	35.9982	46.8858	273_148k.jacobi_data	083_694k.jacobi_data	186_1944k.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl11_refIX2`	40.1429	5.55784	51.3512	21.8283	37.2365	40.3389	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
linear_v2	`bsn1_plvl11_refIX2`	40.2409	3.11203	53.1457	21.3126	37.3002	40.3616	273_148k.jacobi_data	083_694k.jacobi_data	186_1944k.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl12_refIX3`	43.4391	2.69618	57.6641	13.1231	37.3391	44.0014	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
mixed_hot_cold_h32	`bsn1_plvl11_refIX3`	40.1832	17.2861	51.1925	22.5541	33.3143	41.088	284_238k.jacobi_data	259_300k.jacobi_data	285_166k.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl21_refIX3`	42.0505	3.40538	52.0521	16.6528	35.6143	42.7077	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl31_refIX2`	43.7144	1.6942	61.0134	9.06797	39.5274	44.2204	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
linear_v2	`bsn1_plvl51_refIX2`	40.905	2.57378	50.596	17.8659	36.0765	41.5529	273_148k.jacobi_data	083_694k.jacobi_data	186_1944k.jacobi_data
mixed_hot_cold_h32	`bsn1_plvl51_refIX3`	41.4478	15.5725	51.5882	20.1293	31.965	42.7152	284_238k.jacobi_data	259_300k.jacobi_data	285_166k.jacobi_data
mixed_hot_cold_h256	`bsn1_plvl30_refIX3`	46.1283	2.21391	57.852	3.17679	39.2035	47.5668	284_238k.jacobi_data	259_300k.jacobi_data	285_166k.jacobi_data
linear_v2	`bsn1_plvl31_refIX3`	45.818	0.310656	58.7952	2.08586	39.8115	48.1466	273_148k.jacobi_data	083_694k.jacobi_data	241_100k.jacobi_data
mixed_hot_cold_h256	`bsn1_plvl31_refIX3`	45.5994	1.86344	54.5414	2.68721	39.8011	47.1447	284_238k.jacobi_data	259_300k.jacobi_data	285_166k.jacobi_data

Multiple book handling gcc-14 implementations (TOP 20):

orders_table_storage	components	avg_mps	min_mps	max_mps	p90	p50	worst_0	worst_1	worst_2
mixed_hot_cold_h256	`bsn1_plvl30_refIX3`	34.8305	17.4648	39.1389	32.5785	35.2175	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	006_10198k_b38.jacobi_data
mixed_hot_cold_h128	`bsn1_plvl30_refIX3`	34.2601	19.2388	38.7183	31.4695	34.7119	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	020_5859k_b27.jacobi_data
mixed_hot_cold_h256	`bsn1_plvl51_refIX3`	33.4319	22.5239	37.8343	30.4062	33.3048	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	019_8074k_b28.jacobi_data
mixed_hot_cold_h256	`bsn1_plvl11_refIX3`	32.3246	24.4027	35.2288	30.3729	32.5025	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	006_10198k_b38.jacobi_data
mixed_hot_cold_h128	`bsn1_plvl51_refIX3`	33.1125	23.1016	37.6336	29.8193	33.0869	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	019_8074k_b28.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl51_refIX3`	33.3706	21.3732	37.6756	30.8579	33.4887	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	019_8074k_b28.jacobi_data
mixed_hot_cold_h128	`bsn1_plvl11_refIX3`	31.9546	25.0079	34.9747	29.8042	32.0254	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	009_9582k_b34.jacobi_data
mixed_hot_cold_h256	`bsn1_plvl31_refIX3`	34.2458	17.6727	38.7484	31.9541	34.7198	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	006_10198k_b38.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl30_refIX3`	34.6825	15.8842	38.7608	32.5656	35.4017	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	006_10198k_b38.jacobi_data
mixed_hot_cold_h64	`bsn1_plvl30_refIX3`	33.6092	20.6763	38.123	30.5151	33.6862	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	019_8074k_b28.jacobi_data
mixed_hot_cold_h128	`bsn1_plvl31_refIX3`	33.7268	19.7801	38.0169	30.9798	33.9858	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	005_9443k_b38.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl11_refIX3`	32.3343	23.1586	35.1593	30.5123	32.4591	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	006_10198k_b38.jacobi_data
mixed_hot_cold_h64	`bsn1_plvl51_refIX3`	32.674	23.2747	37.1787	29.2151	32.5949	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	019_8074k_b28.jacobi_data
mixed_hot_cold_h64	`bsn1_plvl31_refIX3`	33.1048	21.2774	37.4406	30.098	33.2704	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	006_10198k_b38.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl31_refIX3`	34.1146	15.9556	38.216	32.0533	34.7296	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	006_10198k_b38.jacobi_data
mixed_hot_cold_h64	`bsn1_plvl11_refIX3`	31.3839	24.7721	34.4668	29.1565	31.4075	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	006_10198k_b38.jacobi_data
mixed_hot_cold_h128	`bsn1_plvl12_refIX3`	32.8112	20.214	36.4114	30.3246	33.226	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	020_5859k_b27.jacobi_data
mixed_hot_cold_h256	`bsn1_plvl12_refIX3`	33.2798	18.3924	36.8444	31.2093	33.6557	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	006_10198k_b38.jacobi_data
mixed_hot_cold_h64	`bsn1_plvl12_refIX3`	32.2106	21.4599	35.9702	29.252	32.3979	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	019_8074k_b28.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl12_refIX3`	33.2073	16.6109	36.8834	31.3483	34.0828	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	006_10198k_b38.jacobi_data

Multiple book handling clang-20 implementations (TOP 20):

orders_table_storage	components	avg_mps	min_mps	max_mps	p90	p50	worst_0	worst_1	worst_2	worst_2
mixed_hot_cold_h256	`bsn1_plvl51_refIX3`	37.5125	25.2986	42.0943	34.6049	37.42	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	019_8074k_b28.jacobi_data	186_1944k.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl51_refIX3`	37.5455	24.4102	41.8989	34.942	37.6129	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	019_8074k_b28.jacobi_data	296_353k.jacobi_data
mixed_hot_cold_h256	`bsn1_plvl11_refIX3`	36.4122	27.1037	40.014	34.4293	36.216	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	006_10198k_b38.jacobi_data	186_1944k.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl11_refIX3`	36.5114	26.2557	39.9033	34.478	36.6814	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	006_10198k_b38.jacobi_data	296_353k.jacobi_data
mixed_hot_cold_h128	`bsn1_plvl51_refIX3`	36.9219	25.947	41.8309	33.2855	36.8274	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	020_5859k_b27.jacobi_data	296_353k.jacobi_data
mixed_hot_cold_h128	`bsn1_plvl11_refIX3`	35.8646	27.6775	39.4842	33.6569	35.7045	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	020_5859k_b27.jacobi_data	296_353k.jacobi_data
mixed_hot_cold_h64	`bsn1_plvl51_refIX3`	36.187	25.6996	40.877	32.6276	35.9767	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	019_8074k_b28.jacobi_data	241_100k.jacobi_data
mixed_hot_cold_h256	`bsn1_plvl30_refIX3`	37.6951	19.0124	42.5144	35.3103	37.9512	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	006_10198k_b38.jacobi_data	186_1944k.jacobi_data
mixed_hot_cold_h64	`bsn1_plvl11_refIX3`	35.0682	27.1062	38.7909	32.5907	34.8756	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	019_8074k_b28.jacobi_data	186_1944k.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl30_refIX3`	37.6146	17.6002	42.1854	35.3356	38.1704	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	006_10198k_b38.jacobi_data	296_353k.jacobi_data
mixed_hot_cold_h128	`bsn1_plvl30_refIX3`	36.8688	20.654	41.8316	33.8895	36.9676	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	020_5859k_b27.jacobi_data	186_1944k.jacobi_data
mixed_hot_cold_h256	`bsn1_plvl52_refIX3`	36.9209	20.0766	41.7779	33.732	37.2017	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	019_8074k_b28.jacobi_data	296_353k.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl52_refIX3`	36.9146	18.7805	41.5692	34.5369	37.2413	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	019_8074k_b28.jacobi_data	296_353k.jacobi_data
mixed_hot_cold_h128	`bsn1_plvl52_refIX3`	36.1402	21.4752	41.2002	32.415	36.106	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	020_5859k_b27.jacobi_data	285_166k.jacobi_data
mixed_hot_cold_h64	`bsn1_plvl30_refIX3`	35.8968	21.3775	41.0549	32.3458	36.0247	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	019_8074k_b28.jacobi_data	296_353k.jacobi_data
mixed_hot_cold_h64	`bsn1_plvl52_refIX3`	35.2731	22.0682	40.4705	31.4195	35.2188	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	019_8074k_b28.jacobi_data	296_353k.jacobi_data
mixed_hot_cold_h32	`bsn1_plvl51_refIX3`	34.9294	23.7986	39.5401	30.4365	34.7335	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	019_8074k_b28.jacobi_data	186_1944k.jacobi_data
mixed_hot_cold_h256	`bsn1_plvl12_refIX3`	34.8503	20.9191	38.7959	32.4915	35.2412	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	006_10198k_b38.jacobi_data	285_166k.jacobi_data
mixed_hot_cold_h32	`bsn1_plvl11_refIX3`	33.8054	24.8292	37.3119	30.9172	33.7307	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	019_8074k_b28.jacobi_data	285_166k.jacobi_data
mixed_hot_cold_h1024	`bsn1_plvl12_refIX3`	34.8447	19.7437	38.6258	32.5715	35.4429	016_18100k_b33.jacobi_data	017_10014k_b31.jacobi_data	006_10198k_b38.jacobi_data	241_100k.jacobi_data

Single book handling overall (TOP 20):

compiler	orders_table_storage	components	avg_mps	min_mps	max_mps	p99	p90	p50	worst_0	worst_1	worst_2
clang20	linear_v2	`bsn1_plvl11_refIX3`	44.3905	3.08987	58.457	22.3295	38.4883	44.9381	273_148k.jacobi_data	083_694k.jacobi_data	186_1944k.jacobi_data
clang20	mixed_hot_cold_h1024	`bsn1_plvl11_refIX3`	43.9256	5.56357	54.0438	21.6776	38.3558	44.6825	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
clang20	linear_v2	`bsn1_plvl51_refIX3`	45.1507	2.55143	54.4846	19.4928	36.7451	46.3802	273_148k.jacobi_data	083_694k.jacobi_data	186_1944k.jacobi_data
clang20	mixed_hot_cold_h1024	`bsn1_plvl51_refIX3`	44.8569	5.05421	54.9282	18.6164	36.1017	46.0817	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
clang20	mixed_hot_cold_h1024	`bsn1_plvl30_refIX3`	46.8915	1.9287	58.5714	10.617	39.6212	47.8927	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
clang20	mixed_hot_cold_h1024	`bsn1_plvl31_refIX3`	46.1905	1.66626	55.1726	8.9965	41.2374	47.4413	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
clang20	linear_v2	`bsn1_plvl30_refIX3`	47.6105	0.883425	58.2893	4.83673	41.1537	49.1682	273_148k.jacobi_data	083_694k.jacobi_data	241_100k.jacobi_data
clang20	linear_v1	`bsn1_plvl11_refIX3`	43.9377	3.12486	56.701	13.8634	38.2099	44.6046	273_148k.jacobi_data	083_694k.jacobi_data	186_1944k.jacobi_data
clang20	linear_v1	`bsn1_plvl51_refIX3`	45.3074	2.55738	55.4899	11.6867	35.9982	46.8858	273_148k.jacobi_data	083_694k.jacobi_data	186_1944k.jacobi_data
clang20	mixed_hot_cold_h1024	`bsn1_plvl11_refIX2`	40.1429	5.55784	51.3512	21.8283	37.2365	40.3389	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
clang20	linear_v2	`bsn1_plvl11_refIX2`	40.2409	3.11203	53.1457	21.3126	37.3002	40.3616	273_148k.jacobi_data	083_694k.jacobi_data	186_1944k.jacobi_data
clang20	mixed_hot_cold_h1024	`bsn1_plvl12_refIX3`	43.4391	2.69618	57.6641	13.1231	37.3391	44.0014	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
gcc14	mixed_hot_cold_h1024	`bsn1_plvl30_refIX3`	44.2884	1.88874	55.1051	9.5846	37.9558	45.5046	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
clang20	mixed_hot_cold_h32	`bsn1_plvl11_refIX3`	40.1832	17.2861	51.1925	22.5541	33.3143	41.088	284_238k.jacobi_data	259_300k.jacobi_data	285_166k.jacobi_data
clang20	mixed_hot_cold_h1024	`bsn1_plvl21_refIX3`	42.0505	3.40538	52.0521	16.6528	35.6143	42.7077	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
clang20	mixed_hot_cold_h1024	`bsn1_plvl31_refIX2`	43.7144	1.6942	61.0134	9.06797	39.5274	44.2204	285_166k.jacobi_data	174_880k.jacobi_data	296_353k.jacobi_data
clang20	linear_v2	`bsn1_plvl51_refIX2`	40.905	2.57378	50.596	17.8659	36.0765	41.5529	273_148k.jacobi_data	083_694k.jacobi_data	186_1944k.jacobi_data
clang20	mixed_hot_cold_h32	`bsn1_plvl51_refIX3`	41.4478	15.5725	51.5882	20.1293	31.965	42.7152	284_238k.jacobi_data	259_300k.jacobi_data	285_166k.jacobi_data
clang20	mixed_hot_cold_h256	`bsn1_plvl30_refIX3`	46.1283	2.21391	57.852	3.17679	39.2035	47.5668	284_238k.jacobi_data	259_300k.jacobi_data	285_166k.jacobi_data
clang20	linear_v2	`bsn1_plvl31_refIX3`	45.818	0.310656	58.7952	2.08586	39.8115	48.1466	273_148k.jacobi_data	083_694k.jacobi_data	241_100k.jacobi_data

Multiple book handling overall (TOP 20):

Repeats clang TOP 20, because none of the gcc implementations made it to the top.

Visual Reports

Gcc-14:

Cland-20:

Reflection on the Results

After aggregating benchmark results to SQLite database, several Order Book implementations winners emerged, along with a few pathological edge cases.

Here are key observations coming out of the benchmark results:

The Index Backend. boost::unordered_flat_map (refIX3) is the the best for the throughput. It occupied almost every slot in the Top 20 configurations, consistently beating both tsl::robin_map (refIX2) and Google’s absl::flat_hash_map (refIX4). Even when weighting the results by different criteria, the Boost flat map dominates top list.
Price Level Storage. The highest performing variants were plvl30/31 (Embedded index-based linked list with SOA) and plvl51/52 (Chunked SOA). That said, simple list implementations (plvl11/12) remained surprisingly competitive when compiled with Clang 20.
Orders Table Storage. Best performers: different kinds of mixed_hot_cold and linear_v2/v1
For single book handling all entries in the TOP20 performs very poor in some cases. Which is manifested with a min_mps and p99 values that are 10-30x less performant than an average values.

Edge Cases

Usually average numbers are not the goal when building trading systems. The edge cases are important.

When analyzing the single-book runs, many implementations in TOP 20 exhibited severe performance degradation on specific datasets. In these isolated cases, the minimum throughput (min_mps) and tail latency (P99) dropped 10x to 30x below the median.

Isolating the Pathologies

To figure out exactly which cases (events data files) break the “best” Order Book implementations, I wrote the following query to isolate the runs where the top configurations choked (dropping below 5 Million events/sec):

-- Use one of the table-names defined above:
WITH measurements AS (
    SELECT * FROM all_single
    -- Consider best orders table approaches:
    WHERE orders_table_storage in ( 'mixed_hot_cold_h1024', 'mixed_hot_cold_h256', 'mixed_hot_cold_h128',
                                    'mixed_hot_cold_h64', 'mixed_hot_cold_h32', 'linear_v2', 'linear_v1')

    -- Consider best price level approaches:
    AND ( components like '%plvl3%' OR components like '%plvl5%' OR components like '%plvl1%')

    -- Consider best order references index:
    AND components like '%refIX3'
)

-- Aggregate poor measurements by file
-- and calculate avg performance with those files
SELECT poor_runs.jacobi_data, ROUND(sum(mps)/count(1), 2) as avg_mps, count(1) as cnt,
       SUM( CASE WHEN poor_runs.orders_table_storage like 'linear%' THEN 1 ELSE 0 END ) AS linear_affected,
       SUM( CASE WHEN poor_runs.orders_table_storage like '%hot_cold%' THEN 1 ELSE 0 END ) AS hot_cold_affected
FROM
(
    -- Poor measurements for top implementation (having not more than 5 Mps)
    SELECT compiler, orders_table_storage,
           components, jacobi_data, sum(mps) as mps, count(1) as total_cnt
    FROM measurements
    GROUP BY compiler, orders_table_storage, components, jacobi_data
    HAVING mps < 5 ) poor_runs
GROUP BY poor_runs.jacobi_data
HAVING cnt >=5
ORDER BY avg_mps

jacobi_data	avg_mps	cnt	linear_affected	hot_cold_affected
273_148k.jacobi_data	1.37	24	24	0
083_694k.jacobi_data	2.59	16	16	0
284_238k.jacobi_data	3.03	24	2	22
285_166k.jacobi_data	3.05	24	2	22
259_300k.jacobi_data	3.28	24	2	22
241_100k.jacobi_data	3.37	8	8	0
186_1944k.jacobi_data	3.83	9	9	0

Important takeaway: Look closely at the last two columns. The failures are almost perfectly mutually exclusive.

The “Linear Killers” books affect linear_v1/2 approaches while mixed_hot_cold survives unharmed, and vice versa.

You can verify these pathological datasets yourself. They are available in difficult_dataset_files.tar.bz2.

Here are a few sample commands to start with:

# LINEAR PATHOLOGY REPRODUCTION:
JACOBI_BENCHMARK_EVENTS_FILE=$data_dir/273_148k.jacobi_data $bin_dir/_bench.throughput.linear_v1_levels_storage.single_book --benchmark_min_time=1s "--benchmark_filter=bsn1_plvl[135]._refIX3"
JACOBI_BENCHMARK_EVENTS_FILE=$data_dir/083_694k.jacobi_data $bin_dir/_bench.throughput.linear_v1_levels_storage.single_book --benchmark_min_time=1s "--benchmark_filter=bsn1_plvl[135]._refIX3"
JACOBI_BENCHMARK_EVENTS_FILE=$data_dir/273_148k.jacobi_data $bin_dir/_bench.throughput.linear_v2_levels_storage.single_book --benchmark_min_time=1s "--benchmark_filter=bsn1_plvl[135]._refIX3"
JACOBI_BENCHMARK_EVENTS_FILE=$data_dir/083_694k.jacobi_data $bin_dir/_bench.throughput.linear_v2_levels_storage.single_book --benchmark_min_time=1s "--benchmark_filter=bsn1_plvl[135]._refIX3"

# HOT_COLD PATHOLOGY REPRODUCTION:
JACOBI_BENCHMARK_HOT_STORAGE_SIZE=32 JACOBI_BENCHMARK_EVENTS_FILE=$data_dir/284_238k.jacobi_data $bin_dir/_bench.throughput.mixed_hot_cold_levels_storage.single_book --benchmark_min_time=1s "--benchmark_filter=bsn1_plvl[135]._refIX3"
JACOBI_BENCHMARK_HOT_STORAGE_SIZE=128 JACOBI_BENCHMARK_EVENTS_FILE=$data_dir/284_238k.jacobi_data $bin_dir/_bench.throughput.mixed_hot_cold_levels_storage.single_book --benchmark_min_time=1s "--benchmark_filter=bsn1_plvl[135]._refIX3"
JACOBI_BENCHMARK_HOT_STORAGE_SIZE=256 JACOBI_BENCHMARK_EVENTS_FILE=$data_dir/259_300k.jacobi_data $bin_dir/_bench.throughput.mixed_hot_cold_levels_storage.single_book --benchmark_min_time=1s "--benchmark_filter=bsn1_plvl[135]._refIX3"
JACOBI_BENCHMARK_HOT_STORAGE_SIZE=1024 JACOBI_BENCHMARK_EVENTS_FILE=$data_dir/285_166k.jacobi_data $bin_dir/_bench.throughput.mixed_hot_cold_levels_storage.single_book --benchmark_min_time=1s "--benchmark_filter=bsn1_plvl[135]._refIX3"

Gather an Analise Perf-stats Telemetry

Let’s dive into performance details in worst cases.

To understand why the memory layouts fail on specific files, we will look at the hardware counters using JACOBI_BENCHMARK_PROFILE_MODE=4 (which wraps the execution in perf stat -d ..., for details refer to Benchmarking Suite). We filter for a single implementation and run it for 8 seconds to gather stable telemetry.

Note: performance analysis given below uses a different (less performant) machine but results correlate.

Linear Pathological Case

Inspecting linear_v1 against a “Linear Killer” dataset (273_148k):

JACOBI_BENCHMARK_PROFILE_MODE=4 JACOBI_BENCHMARK_EVENTS_FILE=$data_dir/273_148k.jacobi_data taskset -c 3 $bin_dir/_bench.throughput.linear_v1_levels_storage.single_book --benchmark_min_time=8s "--benchmark_filter=bsn1_plvl11_refIX3"

Perf output:

 Performance counter stats for process id '4059105':

         11,508.26 msec task-clock                #    1.000 CPUs utilized
               133      context-switches          #    0.012 K/sec
                 0      cpu-migrations            #    0.000 K/sec
         5,108,042      page-faults               #    0.444 M/sec
    45,457,074,862      cycles                    #    3.950 GHz                      (49.96%)
    51,568,102,608      instructions              #    1.13  insn per cycle           (62.47%)
     7,506,119,227      branches                  #  652.237 M/sec                    (62.47%)
        47,964,114      branch-misses             #    0.64% of all branches          (62.49%)
    12,741,124,932      L1-dcache-loads           # 1107.129 M/sec                    (62.53%)
       961,678,474      L1-dcache-load-misses     #    7.55% of all L1-dcache hits    (62.55%)
       518,958,071      LLC-loads                 #   45.094 M/sec                    (50.02%)
       262,844,348      LLC-load-misses           #   50.65% of all LL-cache hits     (49.98%)

      11.510932764 seconds time elapsed

This perf-stats clearly shows that memory access is the root cause of poor performance:

# Read "normally" as "Checking the metric on other events files and 10s runs"

# 1. Memory Thrashing
# Extreme number of page-faults
         5,108,042      page-faults               #    0.444 M/sec
# Normally this metric is at most 2-digit number. Here it is measured in millions.

# 2. The Cache Wall
# The other issue high LLC-load-misses,
       262,844,348      LLC-load-misses           #   50.65% of all LL-cache hits     (49.98%)
# Normally it is in range 12-25%:

# 2. Instruction Starvation
# And most likely cause by the above:
    51,568,102,608      instructions              #    1.13  insn per cycle           (62.47%)
# Normally it is 2.0+

The Hypothesis: The market data contains extreme price swings. This forces the underlying linear storage (std::vector<PLVL>) to constantly reallocate and move memory as the top/bottom bounds of the order book expand beyond the current vector capacity.

Confirmation: deeper inspection of the datasets perfectly corroborates this guess. The linear implementations are being destroyed by extreme outlier orders (often called “stub quotes”).

273_148k.jacobi_data: For the first ~75K events, the book operates comfortably in a tight price range of [2100, 10,000]. Then, an extreme SELL order is added at price 1,455,000 (normalized). Because the linear approach maps prices to vector indices, this single event forces the std::vector<PLVL> to immediately allocate ~1.44 million new, empty price levels. This not only wastes massive amounts of memory, but as the price subsequently moves and sets new all-time bounds, the vector is forced to reallocate again. Moving millions of non-trivial price-level objects via move-constructors completely stalls the CPU pipeline.
083_694k.jacobi_data: Exhibits an identical pathology. After ~330K events, an outlier SELL order hits at 1,803,000. The resulting massive reallocation is followed by further price degradation, triggering secondary reallocations.
241_100k.jacobi_data: Follows the exact same pattern.
186_1944k.jacobi_data: Abuses the linear array in the same way (an extreme SELL order arrives at 3,100,000). However, it performs slightly better than the others because the price does not continue to push the bounds outward, sparing the system from subsequent secondary reallocations.

The Mixed Hot/Cold Pathological Case

Sample command to inspect perf-stats (hot storage size is 128 but other sizes are used too):

JACOBI_BENCHMARK_PROFILE_MODE=4 JACOBI_BENCHMARK_HOT_STORAGE_SIZE=128 JACOBI_BENCHMARK_EVENTS_FILE=$data_dir/284_238k.jacobi_data taskset -c 3 $bin_dir/_bench.throughput.mixed_hot_cold_levels_storage.single_book --benchmark_min_time=8s "--benchmark_filter=bsn1_plvl11_refIX3"

Perf output:

 Performance counter stats for process id '4035219':

         11,078.17 msec task-clock                #    1.000 CPUs utilized
               132      context-switches          #    0.012 K/sec
                 0      cpu-migrations            #    0.000 K/sec
                 0      page-faults               #    0.000 K/sec
    44,220,321,575      cycles                    #    3.992 GHz                      (49.99%)
   118,594,302,870      instructions              #    2.68  insn per cycle           (62.53%)
    22,798,601,282      branches                  # 2057.975 M/sec                    (62.53%)
        77,072,117      branch-misses             #    0.34% of all branches          (62.53%)
    41,850,245,460      L1-dcache-loads           # 3777.722 M/sec                    (62.53%)
        39,042,341      L1-dcache-load-misses     #    0.09% of all L1-dcache hits    (62.50%)
         2,097,293      LLC-loads                 #    0.189 M/sec                    (49.96%)
           426,777      LLC-load-misses           #   20.35% of all LL-cache hits     (49.96%)

      11.081143946 seconds time elapsed

This perf-stats actually looks solid: zero page faults, excellent IPC (2.68), and normal cache misses. The CPU is running with good efficiency, but the throughput is still low.

The Hypothesis: The dataset contains a “flickering” outlier price level(s). An isolated top-of-book order appears and vanishes rapidly, far away from the dense cluster of the book. Because the algorithm enforces that the Top Price must stay within the Hot Storage window, this flickering order forces the algorithm to constantly “slide” the hot storage window back and forth across empty price gaps.

When hot storage size - h is small (32), the slide is comparatively cheap.
When h is medium (256), the slide requires heavy bookkeeping, tanking performance.
When h is massive (1024 or 2048), the Hot Storage window is finally wide enough to encompass both the dense cluster and the outlier simultaneously, eliminating the need to slide at all or the ranges of price levels covered by window old and new position start overlap.

That is also backed up by measurements from benchmarks:

SELECT g.jacobi_data, g.orders_table_storage, g.components, g.mps as gcc_mps, c.mps as clang_mps
    FROM gcc14_single g
        INNER JOIN clang20_single c
            ON g.jacobi_data=c.jacobi_data
                AND g.orders_table_storage = c.orders_table_storage
                AND g.components = c.components
WHERE g.orders_table_storage like 'mixed_hot_cold%'
      AND g.components = 'bsn1_plvl11_refIX3'
      AND g.jacobi_data = '284_238k.jacobi_data'
ORDER BY g.id

jacobi_data	orders_table_storage	components	gcc_mps	clang_mps
284_238k.jacobi_data	mixed_hot_cold_h32	bsn1_plvl11_refIX3	16.5721	17.2861
284_238k.jacobi_data	mixed_hot_cold_h64	bsn1_plvl11_refIX3	11.7903	11.8195
284_238k.jacobi_data	mixed_hot_cold_h128	bsn1_plvl11_refIX3	7.97784	7.84732
284_238k.jacobi_data	mixed_hot_cold_h256	bsn1_plvl11_refIX3	6.21393	6.10035
284_238k.jacobi_data	mixed_hot_cold_h1024	bsn1_plvl11_refIX3	20.5684	38.0535

Similar results found for 259_300k.jacobi_data, while for `` performance decreases the higher hot storage size is.

jacobi_data	orders_table_storage	components	gcc_mps	clang_mps
285_166k.jacobi_data	mixed_hot_cold_h32	bsn1_plvl11_refIX3	21.0058	22.5541
285_166k.jacobi_data	mixed_hot_cold_h64	bsn1_plvl11_refIX3	15.9255	16.3427
285_166k.jacobi_data	mixed_hot_cold_h128	bsn1_plvl11_refIX3	11.5171	11.5117
285_166k.jacobi_data	mixed_hot_cold_h256	bsn1_plvl11_refIX3	8.59196	8.46608
285_166k.jacobi_data	mixed_hot_cold_h1024	bsn1_plvl11_refIX3	5.73371	5.56357

But manually checking h=2048 it once again becomes satisfactory.

Confirmation: Upon checking the data we see lots of wide jumps of the top price.

284_238k.jacobi_data

jump size	count
100+	18939
200+	11531
300+	4961
400+	1036
500+	30

284_238k.jacobi_data:

jump size	count
100+	8589
200+	5777
300+	3996
400+	3525
500+	3523

This site is open source. Improve this page.