PCJ is a library for Java language that helps to perform parallel and distributed calculations. The current version is able to work on the multicore systems connected with the typical interconnect such as ethernet or infiniband providing users with the uniform view across nodes.

Download PCJ library (jar file of 29.04.2017 ver. 5.0.6)  Latest (bug fixing release)!

Download PCJ manual (pdf) for PCJ 5 New!

The PCJ library can be used with no cost at BSD license. It requires Java 8 and no additional tools or comilers.  The PCJ library for Java 7 is available in the dowload section.

The source code is available at GitHub: https://github.com/hpdcj/pcj

Version 5.0.6 fixes some bugs occuring with the massive communication.

Version 5.0.3 contains support for Intel KNL chips.

Version 5 introduces asyncPut() and asyncGet() methods; put() and get() methods are now synchronous. There is new handling of shared variables. The code developed for PCJ 4 hast to be modified. For details please reffer to the JavaDoc file.

The usage should be acknowledged by reference to the PCJ web site and/or reference to the papers:

Usage of the PCJ:

Contact: bala@icm.edu.pl faramir@icm.edu.pl

The compilation and execution of the parallel applications especially while using some queueing system or another submission environment is not straightforward. In particular, the information abut the nodes parallel application will be running on is not available in advance or even during job submission but is determined when job starts execution.
 
Most of the systems provide such information through the environment variables and files with the list of nodes used for job execution.
 
The list of nodes, especially while miltiprocessor nodes are present can contain multiple lines with the same names. The multiple entries are used, for example while running MPI application, to start multiple instances of the parallel application on the single node.
 
In the case of PCJ library the execution is simple. The most efficient mechanism is to start single Java Virtual Machine on each node. Within this JVM multiple PCJ threads will be run.
 
While running on multiple node, adequate number of JVMs will be started, using ssh or mpiexec command.
 
Please remember, that PCJ threads running within single JVM will use Java Concurrency Library to synchronize and to communicate. Communication between PCJ threads running within different JVMs will be performed using Java Sockets.
 
In such situation in order to run PCJ application we will use two files:
  • nodes.unique - file containing list of nodes used to run JVMs. In principle this list contains unique names (no duplicated names). This file is used by the mpiexec or other command to start parallel application.
  • nodes.txt - file containing list of nodes used to start PCJ threads.This list may contain duplicated entries showing that on the particular node multiple PCJ threads will be started (within single JVM). The number of PCJ threads used to run application (PCJ.threadsCount()) will be equal to the number of lines (entries) in this file.

Linux Cluster

Example commands which can be run from script or interactive shell:
module load pgi/13.4
module load opempi
mpoexec -hostsfile nodes.unique bash -c 'java -cp pcj.jar PcjExampleHelloWorld'

Linux Cluster with SLURM

The execution is similar to the case of Linux Cluster. However, the proper script submitted to the queue to be prepared.
This file contains defnition of the parameters passed to the queueining system.
The parameters include number of nodes required (nodes=128) and indicate that 1 process per node will be executed (ppn=1).
The execution of java application is preceded by the gathering list of the nodes allocated to the job by the queueing system. The unique list of nodes is then stored in the nodes.unique. Please remember that nodes.unique and nodes.txt can be different.
 
File go.csh:
#!/bin/csh
#PBS -N go
#PBS -l nodes =128: ppn =1
#PBS -l mem =512 mb
#PBS -l walltime =0:10:00
#PBS
 
module load openmpi
 
cat $PBS_NODEFILE > nodes.txt
uniq $PBS_NODEFILE > nodes.unique
mpiexec -hostsfile nodes.unique bash -c 'java -d64 -Xnoclassgc -Xrs -cp pcj .jar PcjExampleHelloWorld'

The job is than executed by submitting it with the qsub command: qsub go.csh

IBM Power 7 (AIX) with Load Leveler

In order to optimize execution on the multinode systems like IBM Power 7, the PCJ application should exclusively use computer nodes. However, the number of applications running on each nodes is 1 which is Java VM. The poe command is
used to invoke Java VM's on the nodes reserved for the execution.

 

#@ job_type = parallel

#@ node = 2

#@ tasks_per_node= 1

#@ queue

 

cat $LOADL_HOSTFILE > nodes.txt

uniq $LOADL_HOSTFILE > nodes.unique

 

poe "java -Xnoclassgc -Xmx6g -cp .:pcj.jar PcjHelloWorld" -hfile nodes.unique -statistic print -bindproc yes -task_affinity core
 
The job is than executed by submitting it with the llsubmit command: llsubmit go_power7.csh
 

Cray XC40 with SLURM

The proper script submitted to the queue to be prepared.
This file (go_xc40.sh) contains defnition of the parameters passed to the queueining system.
 

#!/bin/bash -l                                            
#SBATCH -N 132                                                              # Number of nodes
#SBATCH --ntasks-per-node 48                                       # Numer of tasks per node
#SBATCH --mem 5000                                                     # Required RAM
#SBATCH --time=00:10:00                                               # Required time

#SBATCH -A GES-00                                                        # Account

module load java

srun hostname > nodes.txt
srun -N 132 -n 132 -c 48 java -cp .:PCJ-4.1.0.jar PcjHelloWorld        

The job is than executed by submitting it with the command: sbatch go_xc40.sh

The output is in the file slurm-1234.out (1234 is number of slurm job).

Example output looks as follows:

 

PCJ version 4.1.0.SNAPSHOT-8f78778 built on 2016-01-09 12:26:58.311 CET.
Starting PcjHelloWorld with 6336 thread(s)...
Hello from thread 0 out of 6336
Hello from thread 2 out of 6336
Hello from thread 1 out of 6336
Hello from thread 247 out of 6336
 

....