java - Commons Lang StringUtils.replace performance vs String.replace

Question

When I compared performance of Apache's StringUtils.replace() vs String.replace() I was surprised to know that the former is about 4 times faster. I used Google's Caliper framework to measure performance. Here's my test

public class Performance extends SimpleBenchmark {
    String s = "111222111222";

    public int timeM1(int n) {
        int res = 0;
        for (int x = 0; x < n; x++) {
            res += s.replace("111", "333").length();
        }
        return res;
    }

    public int timeM2(int n) {
        int res = 0;
        for (int x = 0; x < n; x++) {
            res += StringUtils.replace(s, "111", "333", -1).length();
        }
        return res;
    }

    public static void main(String... args) {
        Runner.main(Performance.class, args);
    }
}

output

 0% Scenario{vm=java, trial=0, benchmark=M1} 9820,93 ns; ?=1053,91 ns @ 10 trials
50% Scenario{vm=java, trial=0, benchmark=M2} 2594,67 ns; ?=58,12 ns @ 10 trials

benchmark   us linear runtime
       M1 9,82 ==============================
       M2 2,59 =======

Why is that? Both methods seem to do the same work, StringUtils.replace() is even more flexible.

Answer 1

In modern Java, this is not the case anymore. String.replace was improved in Java-9 moving from regular expression to StringBuilder, and was improved even more in Java-13 moving to direct allocation of the target byte[] array calculating its exact size in advance. Thanks to internal JDK features used, like the ability to allocate an uninitialized array, ability to access String coder and ability to use private String constructor which avoids copying, it's unlikely that current implementation can be beaten by a third-party implementation.

Here are my benchmarking results for your test using JDK 8, JDK 9 and JDK 13 (caliper:0.5-rc1; commons-lang3:3.9)

Java 8 (4x slower indeed):

 0% Scenario{vm=java, trial=0, benchmark=M1} 291.42 ns; σ=6.56 ns @ 10 trials
50% Scenario{vm=java, trial=0, benchmark=M2} 70.34 ns; σ=0.15 ns @ 3 trials

benchmark    ns linear runtime
       M1 291.4 ==============================
       M2  70.3 =======

Java 9 (almost equal performance):

 0% Scenario{vm=java, trial=0, benchmark=M2} 99,15 ns; σ=8,34 ns @ 10 trials
50% Scenario{vm=java, trial=0, benchmark=M1} 103,43 ns; σ=9,01 ns @ 10 trials

benchmark    ns linear runtime
       M2  99,1 ============================
       M1 103,4 ==============================

Java 13 (standard method is 38% faster):

 0% Scenario{vm=java, trial=0, benchmark=M2} 91,64 ns; σ=5,12 ns @ 10 trials
50% Scenario{vm=java, trial=0, benchmark=M1} 57,38 ns; σ=2,51 ns @ 10 trials

benchmark   ns linear runtime
       M2 91,6 ==============================
       M1 57,4 ==================