Regarding Java 11 and above, the accuracy issue of ObjectSampler in --total mode

[yanglong1010]

hi Andrei

In this code, the number of calls to testSmallAlloc and testLargeAlloc is different, but their total memory allocations in bytes are close, which can also be proven from the actual running logs.

public class AllocTest {
    private static volatile long n1 = 0;
    private static volatile long n2 = 0;
    private static Thread t1;

    public static void main(String[] args) throws Exception {
        testAlloc();
        while (true) {
            Thread.sleep(1000);
            System.out.println("small total: " + n1);
            System.out.println("large total: " + n2);
            System.out.println();
        }
    }

    private static void testAlloc() {
        t1 = new Thread(() -> {
            while (true) {
                for (int i = 0; i < 256; i++) {
                    n1 += testSmallAlloc(1024);
                }
                n2 += testLargeAlloc(1024 * 256);
            }
        });
        t1.start();
    }

    private static int testSmallAlloc(int size) {
        return realAlloc(size);
    }

    private static int testLargeAlloc(int size) {
        return realAlloc(size);
    }

    private static int realAlloc(int size) {
        byte[] arr = new byte[size];
        return arr.length;
    }
}

java11/bin/java -Xint AllocTest

small total: 2059834368
large total: 2060451840

small total: 4829180928
large total: 4829216768

small total: 7630267392
large total: 7630225408

small total: 10429923328
large total: 10429923328

Using --total mode, it is expected that the proportions of testSmallAlloc and testLargeAlloc should be close in the flame graph, but in fact testLargeAlloc is 99.91% while testSmallAlloc is only 0.81%.

/root/async-profiler-fork/build/bin/asprof --alloc 256k -f static/profile.html --total -d 10 jps

The reason is that size is used as counter in ObjectSampler::recordAllocation, which causes only the array size to be recorded during testSmallAlloc, which is actually the interval in JVMTI_EVENT_SAMPLED_OBJECT_ALLOC (here we ignore the randomness of interval in JVMTI).

async-profiler/src/objectSampler.cpp

Line 163 in 915b090

Profiler::instance()->recordSample(NULL, size, event_type, &event);

I feel that event._total_size should be used as counter here. I verified it on my branch and it looks much normal after the modification.

if you think this optimization is feasible, I can create a PR.

[apangin]

Thank you for the analysis. Your conclusion makes total sense. Feel free to submit a PR.

[zdyj3170101136]

i also noticed that the small_alloc frame is absent in --ctack=dwarf, but in --cstack=vm mode.

./asprof -d 30 -e cpu -o flamegraph -f cpu_vm.html --cstack vm 632756

./asprof -d 30 -e cpu -o flamegraph -f cpu_dwarf.html --cstack dwarf 632756