Lets you bind a thread to a given core, this can improve performance (this library works best on linux).
OpenHFT Java Thread Affinity library
See affinity/src/test/java for working examples of how to use this library.
The library detects the running platform in Affinity.java
and selects an
implementation for that OS. Features differ between systems:
-
Linux - full affinity control via JNA. The implementation can get and set thread affinity, query the current CPU, and obtain process and thread IDs.
-
Windows - thread affinity is managed through the kernel API. Process and thread IDs are available, while
getCpu()
returns-1
. -
macOS - provides process and thread IDs but does not modify affinity and reports the CPU id as
-1
. -
Solaris - mirrors the macOS implementation: only process and thread IDs are returned with no affinity or CPU querying support.
-
V3.2.0 - Add support for text configuration
-
V3.1.1 - Upgraded JNA dependency to 4.4.0
-
V2.0.1 - Added getThreadId for the process if of the thread.
Java-Thread-Affinity will try to use JNA to provide access to native thread-handling functions. JNA should be installed on your system to get the most from this library.
Java-Thread-Affinity currently depends on JNA version 4.4.0, which in turn depends on a version of GLIBC >= 2.14. If your operating system is an old one, with a version of GLIBC released before 2011, this library will not be able to invoke native functions.
To work around this problem, fork the repository, and override the <version>
tag
for the artifacts jna
and jna-platform
in the project’s pom
file.
choco install jna
Or download jna.jar and jna-platform.jar from the JNA project and add them to your classpath.
The library will read your /proc/cpuinfo
if you have one or provide one and it will determine your CPU layout.
If you don’t have one it will assume every CPU is on one CPU socket.
The library looks for isolated CPUs determined by looking at the CPUs you are not running on by default. i.e. if you have 16 CPUs but 8 of them are not available for general use (as determined by the affinity of the process on startup) it will start assigning to those CPUs.
Note: if you have more than one process using this library you need to specify which CPUs the process can use otherwise it will assign the same CPUs to both processes.
To control which CPUs a process can use, add -Daffinity.reserved={cpu-mask-in-hex}
to the command line of the process.
The mask is a hexadecimal bit mask without
the 0x
prefix where bit 0
represents CPU 0
, bit 1
represents CPU 1
and so on.
Multiple CPUs can be specified by setting more than one bit.
For example:
-
-Daffinity.reserved=2
reserves only CPU1
. -
-Daffinity.reserved=6
reserves CPUs1
and2
. -
-Daffinity.reserved=10
reserves CPUs1
and3
(hexadecimala
).
Use an appropriate mask when starting each process to avoid reserving the same cores for multiple JVMs.
Note: the CPU 0 is reserved for the Operating System, it has to run somewhere.
Java-Thread-Affinity requires that you first isolate some CPU’s.
Once a CPU core is isolated, the Linux scheduler will not use the CPU core to run any user-space processes. The isolated CPUs will not participate in load balancing, and will not have tasks running on them unless explicitly assigned.
To isolate the 1st and 3rd CPU cores (CPU numbers start from 0) on your system, add the following to the kernel command line during boot:
isolcpus=1,3
Using GRUB
sudo sed -i 's/^GRUB_CMDLINE_LINUX_DEFAULT="/GRUB_CMDLINE_LINUX_DEFAULT="isolcpus=1,3 /' /etc/default/grub
sudo update-grub
sudo reboot
Using systemd-boot
sudo sed -i 's/^options \(.*\)/options \1 isolcpus=1,3/' /boot/loader/entries/*.conf
sudo reboot
You can acquire a lock for a CPU in the following way:
In Java 6
AffinityLock al = AffinityLock.acquireLock();
try {
// do some work locked to a CPU.
} finally {
al.release();
}
In Java 7 or 8
try (AffinityLock al = AffinityLock.acquireLock()) {
// do some work while locked to a CPU.
}
You have further options such as
You can reserve a whole core. If you have hyper-threading enabled, this will use one CPU and leave it’s twin CPU unused.
try (AffinityLock al = AffinityLock.acquireCore()) {
// do some work while locked to a CPU.
}
You can chose a layout relative to an existing lock.
try (final AffinityLock al = AffinityLock.acquireLock()) {
System.out.println("Main locked");
Thread t = new Thread(new Runnable() {
@Override
public void run() {
try (AffinityLock al2 = al.acquireLock(AffinityStrategies.SAME_SOCKET,
AffinityStrategies.ANY)) {
System.out.println("Thread-0 locked");
}
}
});
t.start();
}
In this example, the library will prefer a free CPU on the same Socket as the first thread, otherwise it will pick any free CPU.
The AffinityStrategies
enum defines hints for selecting a CPU relative to an existing lock.
Strategy | Meaning |
---|---|
|
Use any available CPU. |
|
Select a CPU on the same core. |
|
Select a CPU on the same socket but a different core. |
|
Select a CPU on another core (possibly another socket). |
|
Select a CPU on a different socket. |
You can get the current thread id using
int threadId = AffinitySupport.getThreadId();
You can get the current CPU being used by
int cpuId = AffinitySupport.getCpu();
The affinity of the process on start up is
long baseAffinity = AffinityLock.BASE_AFFINITY;
The available CPU for reservation is
long reservedAffinity = AffinityLock.RESERVED_AFFINITY;
If you want to get/set the affinity directly you can do
long currentAffinity = AffinitySupport.getAffinity();
AffinitySupport.setAffinity(1L << 5); // lock to CPU 5.
Several examples print the current CPU assignments using AffinityLock.dumpLocks()
.
Each line of the output begins with the zero based CPU id followed by the status
of that CPU. Example output might look like:
0: Reserved for this application
1: Thread[reader,5,main] alive=true
2: General use CPU
3: CPU not available
The number on each line is the logical CPU index as recognised by the library.
The text after the colon describes whether that CPU is free, reserved or already
bound to a thread.
Use these indices when calling AffinityLock.acquireLock(n)
or when constructing explicit affinity masks.
AffinityLock stores a small lock file for each CPU. These files are placed in
the directory specified by the java.io.tmpdir
system property, which by
default points to your system’s temporary directory (usually /tmp
on Linux).
If you want to keep the lock files elsewhere, set this property before using any affinity APIs:
java -Djava.io.tmpdir=/path/to/dir ...
or in code
System.setProperty("java.io.tmpdir", "/path/to/dir");
For a detailed of view of the current affinity state (as seen by the library), execute the following script on Linux systems:
# change to the affinity lock-file directory (defaults to system property java.io.tmpdir)
$ cd /tmp
# dump affinity state
$ for i in "$(ls cpu-*)";
do PID="$(cat $i | head -n1)"; TIMESTAMP="$(cat $i | tail -n1)";
echo "pid $PID locked at $TIMESTAMP in $i"; taskset -cp $PID;
cat "/proc/$PID/cmdline"; echo; echo
done
pid 14584 locked at 2017.10.30 at 10:33:24 GMT in cpu-3.lock
pid 14584's current affinity list: 3
/opt/jdk1.8.0_141/bin/java ...
AffinityThreadFactory
binds each thread it creates according to a set of AffinityStrategy
rules.
This allows executors to automatically run tasks on
cores selected by the library.
ExecutorService es = Executors.newFixedThreadPool(4,
new AffinityThreadFactory("worker",
AffinityStrategies.SAME_CORE,
AffinityStrategies.DIFFERENT_SOCKET,
AffinityStrategies.ANY));
es.submit(() -> {
// your task here
});
System.out.println("\nThe assignment of CPUs is\n" + AffinityLock.dumpLocks());
For an article on how much difference affinity can make and how to use it http://vanillajava.blogspot.com/2013/07/micro-jitter-busy-waiting-and-binding.html
I am currently working on a project related to deadlock detection in multithreaded programs in java. We are trying to run threads on different processors and thus came across your github posts regarding the same. https://github.com/peter-lawrey/Java-Thread-Affinity/wiki/Getting-started Being a beginner, I have little knowledge and thus need your assistance. We need to know how to run threads on specified cpu number and then switch threads when one is waiting.
// lock a cpuId
try (AffinityLock lock = AffinityLock.acquireLock(n)) {
}
where n is the cpu you want to run the thread on.
OR
// lock one of the last CPUs
try (AffinityLock lock = AffinityLock.acquireLockLastMinus(n)) {
}
I have the cpuId in a configuration file, how can I set it using a string?
try (AffinityLock lock = AffinityLock.acquireLock("last")) {
assertEquals(PROCESSORS - 1, Affinity.getCpu());
}
try (AffinityLock lock = AffinityLock.acquireLock("last-1")) {
assertEquals(PROCESSORS - 2, Affinity.getCpu());
}
try (AffinityLock lock = AffinityLock.acquireLock("1")) {
assertEquals(1, Affinity.getCpu());
}
try (AffinityLock lock = AffinityLock.acquireLock("any")) {
assertTrue(lock.bound);
}
try (AffinityLock lock = AffinityLock.acquireLock("none")) {
assertFalse(lock.bound);
}
try (AffinityLock lock = AffinityLock.acquireLock((String) null)) {
assertFalse(lock.bound);
}
try (AffinityLock lock = AffinityLock.acquireLock("0")) { // prints a warning
assertFalse(lock.bound);
}