引言:从“经验值”到无限大
我对数值计算的执念,来自初中时代烟雾缭绕的网吧。那时玩《伝奇》,最让我着迷的不是打怪爆装备,而是角色面板里那条长长的经验值。看着数字不断跳动、累积,最终“叮”一声升级,那种简单的数值驱动整个世界运转的感觉,实在太奇妙了。
当时自学编程,从 G-BASIC 里只有 16 位的 INTEGER,到第一次发现 QBASIC LONG 能存下“20亿”时的兴奋,再到如今成为一名 .NET 程序员,手握理论上“无限大”的 System.Numerics.BigInteger。我对大数的迷恋从未改变,只是疑惑随之而来:
作为 .NET 开发者,我手里的 BigInteger 到底够不够快?特别是和隔壁 Java 的 BigInteger 相比,究竟谁更胜一筹?
尤其是涉及高精度计算、密码学等关键领域时,这不仅是好奇,更是关乎性能的严肃拷问。今天,我就通过一次详尽的对比测试(含 .NET 10、Java 21以及Java 8),来为大家揭晓答案。结果可能出乎意料,请务必看到最后,文末还有一个高性能的“彩蛋”。
实验目标与范围
我们对比这三套实现:
- .NET:System.Numerics.BigInteger(基于 .NET 10)
- Java:java.math.BigInteger
- Java 21(最新LTS):代表未来趋势
- Java 8(国内存量最大):代表庞大的现状
测试操作(3类,覆盖常见大数热点)
- ADD_MOD:(a + b) mod m(加法 + 取模)
- MUL_MOD:(a * b) mod m(乘法 + 取模)
- MODPOW:a^e mod m(模幂,密码学等场景的性能热点)
位宽(3档)
计时口径(尽量抑制噪声)
- 热身:预热 5 次,让 JIT(即时编译器)充分发挥。
- 测量:正式跑 11 次,取中位数,避免单次抖动干扰。
- 指标:nsPerOp(每次操作耗时多少纳秒),这个值越小越好。
- 防作弊:用 XOR 聚合每次计算结果,防止聪明的编译器把整个循环优化掉。
测试环境
说明:以下是在同一 Linux 容器内完成。容器有资源限制,因此“看到的 CPU 核数/内存”与宿主机不完全一致。
- OS:Ubuntu 24.04.3 LTS
- CPU(宿主机型号):AMD EPYC 7763 64-Core Processor
容器可用核心数:2 核(受容器限制)
- 内存:2GB(容器限制)
- .NET
- SDK:10.0.101
- Runtime:10.0.1
- 环境变量:DOTNET_GCServer=1
- Java
- Java 21:OpenJDK 21.0.9(LTS)
- Java 8:Temurin(OpenJDK) 1.8.0_472-b08(广泛使用的 8u 系列)
- JVM 参数(两版本一致):
- -Xms1g -Xmx1g(把 Java 堆固定为 1GB,避免堆动态扩张干扰)
- -XX:+UseG1GC
- -XX:+AlwaysPreTouch(尽量减少运行中页分配扰动)
- 一致性约束
- Java 8/Java 21 使用同一份源码(仅使用 Java 8 语法),并且 用 Java 8 的 javac 编译(classfile=52.0)后分别在 Java 8 / Java 21 上运行,从而尽量把差异归因到运行时/JIT/库实现,而不是编译器生成差异。
赛前插曲:.NET BigInteger 真的“不公平”吗?
有人可能想问:
“.NET BigInteger 每次运算都会创建新的对象(不可变),不公平。”
这句话在“大数场景”里基本成立,但Java BigInteger 同样是不可变类型:add/multiply/mod/xor/modPow 都会返回新值,业务代码层面你并没有公开的 in-place API 可以复用内部缓冲。
因此,在本文的“主对比”(.NET vs Java)里,双方都在不可变范式下运行,这点是公平的。
实验一:.NET BigInteger vs Java 21 BigInteger
3.1 完整源代码
Java 源码:BigIntBench.java
兼容 Java 8 语法;同一份代码在 Java 21 下运行。
- import java.math.BigInteger;import java.util.*;public class BigIntBench { // SplitMix64 RNG (deterministic, fast) static final class SplitMix64 { private long x; SplitMix64(long seed) { this.x = seed; } long nextLong() { long z = (x += 0x9E3779B97F4A7C15L); z = (z ^ (z >>> 30)) * 0xBF58476D1CE4E5B9L; z = (z ^ (z >>> 27)) * 0x94D049BB133111EBL; return z ^ (z >>> 31); } void nextBytes(byte[] dst) { int i = 0; while (i < dst.length) { long v = nextLong(); for (int k = 0; k < 8 && i < dst.length; k++) { dst[i++] = (byte)(v >>> (56 - 8*k)); // big-endian stream } } } } static BigInteger[] genBigInts(int bitSize, int count, long seed) { SplitMix64 rng = new SplitMix64(seed); int byteLen = (bitSize + 7) / 8; BigInteger[] arr = new BigInteger[count]; byte[] buf = new byte[byteLen]; int topBit = (bitSize - 1) % 8; int keepBits = topBit + 1; int firstMask = (keepBits == 8) ? 0xFF : ((1 > 30)) * 0xBF58476D1CE4E5B9UL; z = (z ^ (z >> 27)) * 0x94D049BB133111EBUL; return z ^ (z >> 31); } public void NextBytes(byte[] dst) { int i = 0; while (i < dst.Length) { ulong v = NextUInt64(); for (int k = 0; k < 8 && i < dst.Length; k++) { dst[i++] = (byte)(v >> (56 - 8 * k)); // big-endian stream } } }}enum Op { ADD_MOD, MUL_MOD, MODPOW }record Result(string Lang, int Bits, Op Op, long Ops, double NsPerOp, long Checksum){ public string ToJson() => string.Create(CultureInfo.InvariantCulture, $"{{"lang":"{Lang}","bits":{Bits},"op":"{Op}","ops":{Ops},"nsPerOp":{NsPerOp:F3},"checksum":{Checksum}}}");}static class BigIntBench{ static BigInteger[] GenBigInts(int bitSize, int count, ulong seed) { var rng = new SplitMix64(seed); int byteLen = (bitSize + 7) / 8; var arr = new BigInteger[count]; var buf = new byte[byteLen]; int topBit = (bitSize - 1) % 8; byte topMask = (byte)(1 2_000_000L, 1024 => 1_000_000L, _ => 200_000L }; long mulOps = bits switch { 256 => 500_000L, 1024 => 120_000L, _ => 20_000L }; long powOps = bits switch { 256 => 8_000L, 1024 => 1_500L, _ => 250L }; results.Add(Bench("csharp", bits, Op.ADD_MOD, a, b, null!, mod, addOps, warmups, measures)); results.Add(Bench("csharp", bits, Op.MUL_MOD, a, b, null!, mod, mulOps, warmups, measures)); results.Add(Bench("csharp", bits, Op.MODPOW, ap, null!, ep, mod, powOps, warmups, measures)); } if (json) { foreach (var r in results) Console.WriteLine(r.ToJson()); } else { PrintHuman(results); } return 0; }}
复制代码 3.2 运行方式(可复现命令)
- # Java 21cd /app/bigintbench/javajavac BigIntBench.javajava -Xms1g -Xmx1g -XX:+UseG1GC -XX:+AlwaysPreTouch BigIntBench --json# .NET 10cd /app/bigintbench/csharpdotnet build -c ReleaseDOTNET_GCServer=1 dotnet run -c Release -- --json
复制代码 3.3 实验一原始输出(JSONL)
Java 21:results_java21.jsonl
- {"lang":"java","bits":256,"op":"ADD_MOD","ops":1999872,"nsPerOp":139.589,"checksum":0}{"lang":"java","bits":256,"op":"MUL_MOD","ops":499712,"nsPerOp":387.031,"checksum":0}{"lang":"java","bits":256,"op":"MODPOW","ops":7936,"nsPerOp":17764.884,"checksum":0}{"lang":"java","bits":1024,"op":"ADD_MOD","ops":999424,"nsPerOp":284.672,"checksum":0}{"lang":"java","bits":1024,"op":"MUL_MOD","ops":119808,"nsPerOp":3094.540,"checksum":0}{"lang":"java","bits":1024,"op":"MODPOW","ops":1280,"nsPerOp":264577.852,"checksum":0}{"lang":"java","bits":4096,"op":"ADD_MOD","ops":199680,"nsPerOp":900.735,"checksum":0}{"lang":"java","bits":4096,"op":"MUL_MOD","ops":19456,"nsPerOp":32062.554,"checksum":0}{"lang":"java","bits":4096,"op":"MODPOW","ops":256,"nsPerOp":3422756.113,"checksum":0}
复制代码 .NET 10:results_csharp.jsonl
- {"lang":"csharp","bits":256,"op":"ADD_MOD","ops":1999872,"nsPerOp":146.261,"checksum":0}{"lang":"csharp","bits":256,"op":"MUL_MOD","ops":499712,"nsPerOp":560.246,"checksum":0}{"lang":"csharp","bits":256,"op":"MODPOW","ops":7936,"nsPerOp":169713.608,"checksum":0}{"lang":"csharp","bits":1024,"op":"ADD_MOD","ops":999424,"nsPerOp":297.335,"checksum":0}{"lang":"csharp","bits":1024,"op":"MUL_MOD","ops":119808,"nsPerOp":4792.760,"checksum":0}{"lang":"csharp","bits":1024,"op":"MODPOW","ops":1280,"nsPerOp":1938407.720,"checksum":0}{"lang":"csharp","bits":4096,"op":"ADD_MOD","ops":199680,"nsPerOp":1280.760,"checksum":0}{"lang":"csharp","bits":4096,"op":"MUL_MOD","ops":19456,"nsPerOp":36894.568,"checksum":0}{"lang":"csharp","bits":4096,"op":"MODPOW","ops":256,"nsPerOp":20617970.004,"checksum":0}
复制代码 实验二:加入 Java 8 ——看“现状主流”处于什么位置
这一组的关键点是:同一份 Java 源码用 Java 8 编译,分别在 Java 8 与 Java 21 上运行,尽量避免“编译器产物差异”。
4.1 运行方式
- # 编译(Java 8)/path/to/jdk8/bin/javac BigIntBench.java# 运行(Java 8)/path/to/jdk8/bin/java -Xms1g -Xmx1g -XX:+UseG1GC -XX:+AlwaysPreTouch BigIntBench --json# 运行(Java 21)java -Xms1g -Xmx1g -XX:+UseG1GC -XX:+AlwaysPreTouch BigIntBench --json
复制代码 4.2 实验二原始输出(JSONL)
Java 8:results_java8.jsonl
- {"lang":"java","bits":256,"op":"ADD_MOD","ops":1999872,"nsPerOp":207.335,"checksum":0}{"lang":"java","bits":256,"op":"MUL_MOD","ops":499712,"nsPerOp":468.248,"checksum":0}{"lang":"java","bits":256,"op":"MODPOW","ops":7936,"nsPerOp":17697.113,"checksum":0}{"lang":"java","bits":1024,"op":"ADD_MOD","ops":999424,"nsPerOp":390.906,"checksum":0}{"lang":"java","bits":1024,"op":"MUL_MOD","ops":119808,"nsPerOp":3089.474,"checksum":0}{"lang":"java","bits":1024,"op":"MODPOW","ops":1280,"nsPerOp":277395.652,"checksum":0}{"lang":"java","bits":4096,"op":"ADD_MOD","ops":199680,"nsPerOp":990.708,"checksum":0}{"lang":"java","bits":4096,"op":"MUL_MOD","ops":19456,"nsPerOp":30692.214,"checksum":0}{"lang":"java","bits":4096,"op":"MODPOW","ops":256,"nsPerOp":3468269.539,"checksum":0}
复制代码 可视化与总结
从一个 .NET 程序员的视角来看,这次的测试结果可以说既在情理之中,又有些出乎意料。
- ADD_MOD (加法+取模): 在这个项目上,.NET 和 Java 21 几乎打了个平手,差距微乎其微。可以说,在基础的加法运算上,.NET 表现得相当不错。
- MUL_MOD (乘法+取模): 从这里开始,差距出现了。.NET 明显慢于 Java,性能鸿沟开始变得“肉眼可见”。
- MODPOW (模幂): 这是差距最大的地方。.NET 在这项测试中被 Java 21 拉开了 6到9倍 的差距。对于从事密码学或需要大量大数运算的开发者来说,这是一个非常刺眼的信号。
- Java 8 vs Java 21: 毫无疑问,Java 21 在绝大多数情况下都比老迈的 Java 8 要快。不过有趣的是,在 MUL_MOD 的 1024 和 4096 位测试中,Java 8 居然出现了“反超”的现象。这可能是由于 JIT 策略、算法选择的阈值差异,或是单纯的测量误差。虽然这不影响“Java 21更快”的总体结论,但也提醒我们性能测试的复杂性。
总而言之,这次对决让我们清楚地看到,在复杂的大数运算上,.NET 的 BigInteger 确实还有很长的路要走。
One More Thing:当“外援”登场
在寻找 .NET 大数性能优化方案的过程中,我们自然能想到了业界标杆 GMP ——它是 GNU Multi-Precision Arithmetic Library,很多数学软件/密码学实现都会用它做高性能大整数运算。
我碰巧也为它做了一个 .NET 封装:Sdcb.Arithmetic。
但必须提前声明:让 GMP 作为“外援”加入这场对比,是“不公平”的。原因很简单:
- 语言优势:GMP 是原生 C/汇编,而 .NET 和 Java 是在虚拟机上运行的托管语言。
- 内存策略:GMP 鼓励使用 in-place API,可以直接在原地修改数值,大大减少了内存分配和 GC 压力。而 .NET 和 Java 的 BigInteger 则是不可变对象。
所以,这部分的结果更像是一个“彩蛋”,展示的是:如果你愿意引入原生依赖,并改变编码风格,.NET 的大数性能可以达到怎样的高度。
7.1 客串实验完整源代码(.NET BigInteger + GmpInteger 同场)
下面代码会同时输出两套结果:csharp_bigint 与 csharp_gmp_inplace(仍是 JSONL)。
Program.cs(客串版,完整)- using System;using System.Collections.Generic;using System.Diagnostics;using System.Globalization;using System.Linq;using System.Numerics;using Sdcb.Arithmetic.Gmp;sealed class SplitMix64{ private ulong _x; public SplitMix64(ulong seed) => _x = seed; public ulong NextUInt64() { ulong z = (_x += 0x9E3779B97F4A7C15UL); z = (z ^ (z >> 30)) * 0xBF58476D1CE4E5B9UL; z = (z ^ (z >> 27)) * 0x94D049BB133111EBUL; return z ^ (z >> 31); } public void NextBytes(byte[] dst) { int i = 0; while (i < dst.Length) { ulong v = NextUInt64(); for (int k = 0; k < 8 && i < dst.Length; k++) dst[i++] = (byte)(v >> (56 - 8 * k)); } }}enum Op { ADD_MOD, MUL_MOD, MODPOW }record Result(string Impl, int Bits, Op Op, long Ops, double NsPerOp, long Checksum){ public string ToJson() => string.Create(CultureInfo.InvariantCulture, $"{{"lang":"{Impl}","bits":{Bits},"op":"{Op}","ops":{Ops},"nsPerOp":{NsPerOp:F3},"checksum":{Checksum}}}");}static class BenchUtil{ public static void MaskToBitSize(byte[] buf, int bitSize) { int topBit = (bitSize - 1) % 8; int keepBits = topBit + 1; int firstMask = keepBits == 8 ? 0xFF : ((1 a == "--json"); int warmups = 5; int measures = 11; int[] bitSizes = [256, 1024, 4096]; var results = new List(); foreach (int bits in bitSizes) { long addOps = bits switch { 256 => 2_000_000L, 1024 => 1_000_000L, _ => 200_000L }; long mulOps = bits switch { 256 => 500_000L, 1024 => 120_000L, _ => 20_000L }; long powOps = bits switch { 256 => 8_000L, 1024 => 1_500L, _ => 250L }; // BigInteger data var modB = BigIntegerBench.GenModulus(bits, 0xA1B2C3D4E5F60708UL ^ (uint)bits); var aB = BigIntegerBench.Gen(bits, 1024, 0x1111222233334444UL ^ (uint)bits); var bB = BigIntegerBench.Gen(bits, 1024, 0x9999AAAABBBBCCCCUL ^ (uint)bits); var apB = BigIntegerBench.Gen(bits, 256, 0x13579BDF2468ACE0UL ^ (uint)bits); var epB = BigIntegerBench.Gen(Math.Min(bits, 512), 256, 0x0FEDCBA987654321UL ^ (uint)bits); // GmpInteger data (same seeds/bit sizes) using var modG = GmpIntegerBench.GenModulus(bits, 0xA1B2C3D4E5F60708UL ^ (uint)bits); var aG = GmpIntegerBench.Gen(bits, 1024, 0x1111222233334444UL ^ (uint)bits); var bG = GmpIntegerBench.Gen(bits, 1024, 0x9999AAAABBBBCCCCUL ^ (uint)bits); var apG = GmpIntegerBench.Gen(bits, 256, 0x13579BDF2468ACE0UL ^ (uint)bits); var epG = GmpIntegerBench.Gen(Math.Min(bits, 512), 256, 0x0FEDCBA987654321UL ^ (uint)bits); try { results.Add(BenchBigInteger(bits, Op.ADD_MOD, aB, bB, null!, modB, addOps, warmups, measures)); results.Add(BenchBigInteger(bits, Op.MUL_MOD, aB, bB, null!, modB, mulOps, warmups, measures)); results.Add(BenchBigInteger(bits, Op.MODPOW, apB, null!, epB, modB, powOps, warmups, measures)); results.Add(BenchGmpInteger(bits, Op.ADD_MOD, aG, bG, null!, modG, addOps, warmups, measures)); results.Add(BenchGmpInteger(bits, Op.MUL_MOD, aG, bG, null!, modG, mulOps, warmups, measures)); results.Add(BenchGmpInteger(bits, Op.MODPOW, apG, null!, epG, modG, powOps, warmups, measures)); } finally { GmpIntegerBench.DisposeAll(aG); GmpIntegerBench.DisposeAll(bG); GmpIntegerBench.DisposeAll(apG); GmpIntegerBench.DisposeAll(epG); } } if (json) { foreach (var r in results) Console.WriteLine(r.ToJson()); } return 0; }}public static class Program{ public static int Main(string[] args) => Runner.Run(args);}
复制代码 gmpbench.csproj(客串版项目文件)- Exe net10.0 enable enable
复制代码 7.2 客串实验原始输出(JSONL)
- {"lang":"csharp_bigint","bits":256,"op":"ADD_MOD","ops":1999872,"nsPerOp":146.261,"checksum":0}{"lang":"csharp_bigint","bits":256,"op":"MUL_MOD","ops":499712,"nsPerOp":560.246,"checksum":0}{"lang":"csharp_bigint","bits":256,"op":"MODPOW","ops":7936,"nsPerOp":169713.608,"checksum":0}{"lang":"csharp_gmp_inplace","bits":256,"op":"ADD_MOD","ops":1999872,"nsPerOp":76.644,"checksum":0}{"lang":"csharp_gmp_inplace","bits":256,"op":"MUL_MOD","ops":499712,"nsPerOp":114.690,"checksum":0}{"lang":"csharp_gmp_inplace","bits":256,"op":"MODPOW","ops":7936,"nsPerOp":13931.914,"checksum":0}{"lang":"csharp_bigint","bits":1024,"op":"ADD_MOD","ops":999424,"nsPerOp":297.335,"checksum":0}{"lang":"csharp_bigint","bits":1024,"op":"MUL_MOD","ops":119808,"nsPerOp":4792.760,"checksum":0}{"lang":"csharp_bigint","bits":1024,"op":"MODPOW","ops":1280,"nsPerOp":1938407.720,"checksum":0}{"lang":"csharp_gmp_inplace","bits":1024,"op":"ADD_MOD","ops":999424,"nsPerOp":97.260,"checksum":0}{"lang":"csharp_gmp_inplace","bits":1024,"op":"MUL_MOD","ops":119808,"nsPerOp":562.235,"checksum":0}{"lang":"csharp_gmp_inplace","bits":1024,"op":"MODPOW","ops":1280,"nsPerOp":218715.147,"checksum":0}{"lang":"csharp_bigint","bits":4096,"op":"ADD_MOD","ops":199680,"nsPerOp":1280.760,"checksum":0}{"lang":"csharp_bigint","bits":4096,"op":"MUL_MOD","ops":19456,"nsPerOp":36894.568,"checksum":0}{"lang":"csharp_bigint","bits":4096,"op":"MODPOW","ops":256,"nsPerOp":20617970.004,"checksum":0}{"lang":"csharp_gmp_inplace","bits":4096,"op":"ADD_MOD","ops":199680,"nsPerOp":179.720,"checksum":0}{"lang":"csharp_gmp_inplace","bits":4096,"op":"MUL_MOD","ops":19456,"nsPerOp":5431.441,"checksum":0}{"lang":"csharp_gmp_inplace","bits":4096,"op":"MODPOW","ops":256,"nsPerOp":2662198.492,"checksum":0}
复制代码 7.3 客串可视化(以GMP为基准)
总结与展望
从这次“硬碰硬”的对决中,我们可以清晰地看到:在基础加法上,.NET BigInteger 与 Java 不分伯仲;但在乘法,尤其是模幂运算(对密码学等场景极其重要)上,.NET 目前确实存在明显的短板,大幅落后于 Java。
承认不足是改进的开始。对于绝大多数业务场景,内置的 BigInteger 依然够用且方便。但如果你的应用处于性能敏感区(如加密算法、科学计算),那么也许是时候考虑一些“重武器”了。
这也正是我开发并维护 Sdcb.Arithmetic 的初衷。它通过封装 GMP 等高性能原生库,为 .NET 带来了原地修改(in-place)以及高达数倍的性能提升(如文中实验所示)。如果你对性能有极致追求,或者想看看 .NET 在大数计算上的极限,欢迎去 GitHub 点个 Star ⭐,试一试这个库。
感谢阅读!如果你觉得这两个语言的对比分析有意思,或者对 .NET 高性能编程感兴趣,欢迎在评论区留言交流,也欢迎加入我的 .NET骚操作 QQ群:495782587,我们一起探索更多技术硬核玩法。
来源:程序园用户自行投稿发布,如果侵权,请联系站长删除
免责声明:如果侵犯了您的权益,请联系站长,我们会及时删除侵权内容,谢谢合作! |