I have two desktops. One has an AMD FX CPU with 8 processors and the other is an Intel i5 with 4 processors. Their speed is suitable for updating the values of The Bubble Index for all the various assets. However, if I needed to rerun every time-series for every time window, this would take weeks. Each desktop has a GPU ready to be given work. If these GPUs could help compute The Bubble Index, then the time to generate an entire Bubble Index would be magnitudes smaller.

The Bubble Index algorithm involves curve fitting over a (i, j) 2-Dimensional space of parameters. It searches for the best fit over 342 combinations of parameters. On a time series such as TSLA (Tesla Motor), which contains around 1000 daily prices, the 512 day window will run 488 parallel processes. On the AMD FX with all 8 cores, this would take on the order of 40min. On the Intel around 50min. To calculate the 52, 104, 153, 256, 512 windows for TSLA would take around 1.5 hours.

Now, thanks to OpenCL and JavaCL, running all these windows for TSLA now takes only around 3min on an NVIDIA GT 630. Nothing short of Amazing!

The conversion of the Callables and Futures to a structure suitable for GPU parallel processing seemed easy; but the process proved challenging and enlightening. I had plenty of debugging headaches and late nights. The way I proceeded was as follows: First download the NVIDIA driver for my graphics card. Next install the CUDA toolkit. Note: There are really cool CUDA examples written in C included in the toolkit. Then I downloaded Netbeans with Java. A helpful next step was to access the Maven Archetype application called javacl-simple-tutorial. From this point forward, it was a matter of translating my old code into OpenCL and JavaCL code.

The javacl simple tutorial provides a nice starting point to begin building a GPU application. The book, OpenCL in Action was also helpful in understanding the underlying processes involved in memory transfer and kernel execution. One of the biggest issues I had was the allocation of local arrays on the device. I needed each work-item to have its own allocation of a specific float array. Apparently, JavaCL does not support __local float* array (only __local int* array). I finally came to realize that the proper allocation of local memory vs. host memory provided the answer. Since JavaCL does not support __local float*, I struggled to find a solution. In retrospect the solution is simple. Instead of allocating a __local float* array on the device, it is suitable to allocate a large enough array for all work-items to use a single __global float* array shared by the host and device — each work-item having its own private subsection of this massive array.

There were also some difficulties in getting the new values produced by the GPU to match the old values produced by the CPU. Previously, all values were stored as a double. Since not all OpenCL devices support double-precision floating-point format, I decided to convert the calculations to float. The conversion of all values from double to float did change the values by a small percentage. The differences were AT MOST of magnitude 1.0%. The output from several series were compared and I believe the difference is so small in most cases as to be negligible. The switch to JavaCL is now successful.

Thanks to this new overhaul, the normalization of all The Bubble Indices will proceed much faster. Also, I will be adding many new commodities to the site.