CallForFunding

GPUs and next generation hardware - Google's TPU - are the workhorses of Big Data in the cloud.

netlib-java provides access to hardware-accelerated high performance linear algebra from the Java virtual machine. netlib-java ships as part of Spark's MLlib, and is widely used in Hadoop as other platforms. But it can only continue to support and anticipate new hardware architectures and operating systems with funding.

Need

We already see fragmentation in Big Data platforms, with each new application or platform creating tailored support for restricted algorithms on specific chipsets. It is a tactical necessity for each player to repeat the work of their competitors, with no advantage.

By pooling resources, netlib-java can provide the strategic API for linear algebra operations that benefit from hardware acceleration, building on decades of machine code optimisation from the experts.

Benefit

A funded netlib-java means that when you need hardware acceleration for your new algorithm, it will already be available for you to use. When you upgrade your hardware or operating systems, they will already be supported.

Customer-facing businesses can focus on solving their customer's problems, instead of spending their precious human resource on low-level hardware integration and optimisation issues.

Approach

There are a few areas that require improvement:

1. support for complex numbers
2. support for hardware-specific memory regions and NIO
3. an automated, continuous, release
4. MultiBLAS

Complex Numbers / additional algorithms

Complex numbers can only be supported by returning to the Fortran JVM compiler at the core of netlib-java. The compiler hasn't seen any code changes in nearly ten years and doesn't support a complex number representation that maps onto accelerated hardware. A redesign of the compiler would enable complex number support as well as simplifying the build and supporting additional Fortran algorithms beyond BLAS, LAPACK and ARPACK.

Special Memory / Hardware / NIO Safety

Bleeding edge GPUs allow direct access to their memory as if it were in the CPU memory space. netlib-java could support this with a thin native binding layer and direct memory access via NIO. This, again, requires changes to the Fortran compiler.

netlib-java currently uses a kind of native memory access known as critical access, which can cause long delays of garbage collection that may result in OutOfMemory exceptions. Using NIO (as an alternative to) JVM-managed memory would reduce the risk of such corner-case problems.

Continuous Release / OS Support

Today, to release an update to netlib-java requires three physical computers and several hours of effort. With the advent of docker and cloud platforms, we can do better. There is no reason why we can't provide binaries for any operating system that can be virtualised, with snapshot releases on merges to master. The community then becomes empowered and there is even less need for any of us to write non-competitive tactical algorithms.

Funding

If you are interested in funding this initiative, please get in touch by emailing me at sam.halliday@gmail.com

Please note that I am only interested in serious financial commitments, not individual donations. If your organisation is unable to commit at least $10,000 to the pot, then I'm afraid the numbers just don't work out.

Who Are You?

I am Sam Halliday. I am a chartered mathematician and software engineer, based in London. I am the author of netlib-java.

For more details on high performance linear algebra on the JVM, please watch my talk at Scala eXchange 2014 (follow along with high-res slides).