High Performance Visual Data Analytics

February 4, 2013, Electronic Imaging Conference, Burlingame, CA—Wes Bethel from Lawrence Berkeley Lab explained why high performance visual data analytics is both relevant and difficult. The onset of big data and the advances in computers make some areas easier, but the big problems still need a lot of technology to find solutions.

Visualization has always placed an emphasis on multi-disciplinary interactions and gave the world a term for applied graphing. The ’87 paper on “visualization in scientific computing” set the standards for visualization and the ability to scale up and down to see things that are hard to see. Some of the parameters that visualization enable include time scales, physical scales much larger of smaller than a person, and inaccessible objects and events.

All efforts in visualization involve computers and people in the loop to provide feedback the systems for other representations. The necessary high performance is relevant and difficult to achieve, as more and increasingly complex data are viewed. The 10’s of PB of data in some datasets takes upwards of 30 Pflops to process, so the programming tasks are a major bottleneck.

Knowledge relies on our ability to process information and draw conclusions. Challenges include hidden information and human biases. A confirmation bias makes us want to see results that agree with our hypotheses, so it is important to develop experiments that allow for alternative, and unpleasant, outcomes. In a similar vein, our focus and background affect our perceptions.

For example, scientists looking at climate change view the number and frequency of extreme weather events like hurricanes. Other extreme events are atmospheric rivers that transport water, meso-scale connective systems like thunderstorms, and blocking events like heat waves and drought or cold snaps. Most weather models have scalable modules that allow for increased resolution in areas of interest.

The challenge is to identify events and get data. The model resolutions have dropped from 200 km on a side to 12 km and now generate over 100TB of output for a 26-year model run at 25 km per side window. One problem is that the processing scales as a N4 factor and high frequency events become I/O bound in the input side. Programming the models is very difficult.

Some new toolkits like TECA (LBL’s Toolkit for Extreme Climate Analysis) help in finding the data with features and stitching together candidates, both computer intensive tasks. Detecting features like vorticity greater than some threshold plus a low pressure core plus a warm neighbor is a typical task. So far the models work, and can correlate with history. The resolution matters for these computations, as some older, lower resolution models didn’t correlate with history, implying no chance for predicting the future.

For atmospheric rivers, they have to gather information on a group and regional basis. The shape and size of the regions are important criteria. The procedures have to work on different size and shape areas, more challenging than a uniform grid. One model with one parameter took 4 seconds on 9855 CPUs. A full atmospheric model has over 1k parameters. Climate research needs multi-dimensional visualizations and a better software infrastructure.

Another issue facing the researchers is group think. The goal is to get to consensus, but you still need people to ask the hard questions. To make matters worse, the issue of false consensus comes up when the boss is in the meeting. The best way around this is to break up into small groups to open up dialogs.

An example in another field is laser-plasma acceleration. Wakefield acceleration is like electron surfing. Simulations produce dump-time variations and a need to cross correlate. The limiting factor is the number of particles per time step. In ’08, they had 4 GB per core, but needed 5 GB for the simulation to run to completion. After adding the memory, they were able to show a partial oscillation 3-D pattern. The visual data helped to explain the underlying physics. These calculations required parallel indices and 2-D and 3-D visualizations.

LBL does a lot of work with the linear accelerator. If the electro-magnets are not set up properly, particles escape from the beam. Their halo particle investigations required modeling 1B particles over 600 time steps to show that halo exists in the late stages of the accelerator.

Magnetic reconnection is a process where you break and reconnect magnetic fields. One example is the earth’s magnetic field at the poles where the fields change dramatically. One result is the aurora at the poles. The calculations take 120,000 cores to track 1T particles per time step, generating 30TB of data per time step. Flux ropes are asymmetric distributions near the reconnection hot spots. The changing data in the model over time require parallel I/O to provide an end-to-end view of the requirements. It takes over a day to optimize the data for a run. Hooking up 100k parallel machines is not a trivial task.

The Texas sharpshooter fallacy is where a lot of data points are arbitrarily grouped and the result looks like order. The term comes from the story of a person taking random shots at a barn every day, then painting a bulls eye around a cluster of bullet holes. At some scale anything can be made to look like structure.

Investigating the genetic regulatory network involves image acquisition, segmentation, and registration. Then the data are visualized in physical and data views. For expression in genes, researchers look for overlap and correlation. The data slices and cluster analysis lead to comparative analysis, but the clusters seem to change with time.

Expectational bias figures into combustion research. The model is trying to increase efficiency and the primary parameter is turbulence. This topological-based simulation used feature detection to track and analyze the flame front. At some resolutions, the increase in turbulence decreased efficiency, counter to expectations. This unacceptable response led the researchers to evaluate their assumptions and biases before changing resolution to a smaller scale.

These efforts have shown that there are still a lot of technology challenges as larger and more complex data are involved. The questions they have to answer are also getting more complex and must be accessible on evolving technology platforms. The promise of more computer power over time is getting harder to achieve.

The challenges are also in knowledge. We know more now, but are aware that we don’t know a lot. Human biases and group dynamics get in the way of even carefully targeted queries. How do you get someone to ask the proper questions?

The visualization and analysis of datasets is getting more relevant and difficult. There are always pressures to get an answer as soon as possible, and the answers may be a moving target. There are technical and non-technical challenges to any large scale investigation. There are lots of papers on technology and investigative methods. Pictures are ok, but you also need quantitative analysis and correlations to other work.

There is on-going work on the technology front and visualization is not enough. Finding things in the data is much more than just visualization, as levels of uncertainty and other error factors really matter. The I/O bottlenecks are being addressed by in-situ simulations and visualizations while the data are still in memory. Exploratory issues are related to knowing what to do.