I have a question around getting CPU utilization for a given JNI block. I'm making some intensive CPU computation in the underlying C++ JNI native method. I'm in the process of optimizing this computation and want to benchmark it against varying inputs.
I need some guidance on how to go about measuring this. The alternatives I have considered so far are
Using JMX ThreadMXBean
to measure system CPU usage for the current thread that invokes call into JNI method. However, I am not sure if JNI code is executed within the invoking thread context. What happens when the thread spawns more threads?
Using JMX OperatingSystemMXBean
to get the CPU usage for the entire JVM. Ideally, this is not want I want as there could be parallel executions in JVM that might tweak the benchmarking.
Measure externally using getrusage(..)
. What I want to know here is that how is it different than using OperatingSystemMXBean
.
You can use ThreadMXBean
to get cpu usage statistics from all running threads. In the example below the CPU usage per thread
is calculated:
private int sampleTime = 10000;
private ThreadMXBean threadMxBean = ManagementFactory.getThreadMXBean();
private RuntimeMXBean runtimeMxBean = ManagementFactory.getRuntimeMXBean();
private OperatingSystemMXBean osMxBean = ManagementFactory.getOperatingSystemMXBean();
private Map<Long, Long> threadInitialCPU = new HashMap<Long, Long>();
private Map<Long, Float> threadCPUUsage = new HashMap<Long, Float>();
private long initialUptime = runtimeMxBean.getUptime();
ThreadInfo[] threadInfos = threadMxBean.dumpAllThreads(false, false);
for (ThreadInfo info : threadInfos) {
threadInitialCPU.put(info.getThreadId(), threadMxBean.getThreadCpuTime(info.getThreadId()));
}
try {Thread.sleep(sampleTime);} catch (InterruptedException e) {}
long upTime = runtimeMxBean.getUptime();
Map<Long, Long> threadCurrentCPU = new HashMap<Long, Long>();
ThreadInfo[] threadInfos = threadMxBean.dumpAllThreads(false, false);
for (ThreadInfo info : threadInfos) {
threadCurrentCPU.put(info.getThreadId(), threadMxBean.getThreadCpuTime(info.getThreadId()));
}
// CPU over all processes
//int nrCPUs = osMxBean.getAvailableProcessors();
// total CPU: CPU % can be more than 100% (devided over multiple cpus)
long nrCPUs = 1;
// elapsedTime is in ms.
long elapsedTime = (upTime - initialUptime);
for (ThreadInfo info : threadInfos) {
// elapsedCpu is in ns
Long initialCPU = threadInitialCPU.get(info.getThreadId());
if (initialCPU != null) {
long elapsedCpu = threadCurrentCPU.get(info.getThreadId()) - initialCPU;
float cpuUsage = elapsedCpu / (elapsedTime * 1000000F * nrCPUs);
threadCPUUsage.put(info.getThreadId(), cpuUsage);
}
}
// threadCPUUsage contains cpu % per thread
System.out.println(threadCPUUsage);
// You can use osMxBean.getThreadInfo(theadId) to get information on every thread reported in threadCPUUsage and analyze the most CPU intentive threads
On Linux, this method may help you:
- Run
kill -3 <pid of Java process>
to get the thread dump. Thread dump will be added standard output.
- Open the dump and search for the thread you're interested in. (You can search using class name)
- Get the thread PID (it's in hex) and convert it to decimal.
- Run
top
command.
- Press
Shift-H
to enable Threads View
- Identify the thread using decimal value you obtained above.
You can see cpu / memory usage of your thread there.