OpenMP/Java
Third party funded individual grant
Start date :
01.10.2009
End date :
01.10.2015
Project details
Scientific Abstract
JaMP is an implementation of the well-known OpenMP standard adapted for Java. JaMP allows one to program, for example, a parallel for loop or a barrier without resorting to low-level thread programming. For example:
class Test {
...void foo() {
......//#omp parallel for
......for (int i=0;i .........a[i] = b[i] + c[i]
......}
...}
}
is valid JaMP code. JaMP currently supports all of OpenMP 2.0 with partial support for 3.0 features, e.g., the collapse clause. JaMP generates pure Java 1.5 code that runs on every JVM. It also translates parallel for loops to CUDA-enabled graphics cards for extra speed gains. If a particular loop is not CUDA-able, it is translated to a threaded version that uses the cores of a typical multi-core machine. JaMP also supports the use of multiple machines and compute accelerators to solve a single problem. This is achieved by means of two abstraction layers. The lower layer provides abstract compute devices that wrap around the actual CUDA GPUs, OpenCL GPUs, or multicore CPUs, wherever they might be in a cluster. The upper layer provides partitioned and replicated arrays. A partitioned array automatically partitions itself over the abstract compute devices and takes the individual accelerator speeds into account to achieve an equitable distribution. The JaMP compiler applies code-analysis to decide which type of abstract array to use for a specific Java array in the user’s program.
In 2010, the JaMP environment was extended to support the use of multiple machines and compute accelerators to solve a single problem. We have developed two new abstraction layers. The lower layer provides abstract compute devices which wrap around the actual CUDA GPUs, OpenCL GPUs, or multicore CPUs, wherever they might be in a cluster. The upper provides partitioned and replicated arrays. A partitioned array automatically partitions itself over the abstract compute devices and takes the individual accelerator speeds into account to achieve an equitable distribution. The JaMP compiler applies code-analysis to decide which type of abstract array to use for a specific Java array in the user’s program.
In 2012, we extended the JaMP framework to also handle Java objects on multiple ma- chines and accelerators (and not just arrays of primitive types). We added two different ways to handle objects. Standard shared objects are replicated on all compute devices. Arrays of objects are now also replicated or partitioned over the different devices. To increase the performance of the program, the framework has to break with Java’s semantics. Java’s object structure is mapped to a flat memory structure for the execution on the different devices.
In 2013, we examined how to better support Java objects in OpenMP parallel code, regardless of where the code is executed. We found that we needed to restrict the language slightly by forbidding inheritance of objects used in a parallel block. This ensures that the objects will not be of a different type than what is seen at compile time. We use this property to, for example, allow object inlining into arrays to occur naturally. With the added inlining, communication of objects and arrays over the network and to the compute devices was accelerated enormously, including a small performance increase on the devices themselves.
In 2014 we developed a JaMP implementation for Android 4.0. Currently this version only supports the SIMD construct of OpenMP.
In 2015 we added OpenMP tasks (OpenMP 3.0) to JaMP. This makes it possible to parallelize recursive algorithms with JaMP.
Involved:
Contributing FAU Organisations:
Funding Source
