In studying computer science, I always wished for more emphasis on the history of the field. I like to read these formative papers in order to (a) understand the intellectual development of computer science, and (b) inspire new avenues of thinking in future research. Here is my Great Works in Computer Science reading list. Please suggest any additions to the list.
Hydra is a distributed data processing and storage system originally developed at AddThis. It ingests streams of data (think log files) and builds trees that are aggregates, summaries, or transformations of the data. These trees can be used by humans to explore (tiny queries), as part of a machine learning pipeline (big queries), or to support live consoles on websites (lots of queries). Hydra is most similar in feature set to Google Dremel / Apache Drill and Apache Samza.
ssync is a divide-and-conquer file copying tool to multiple destination hosts. It transfers to N remote machines in log N iterations. ssync is a thin transparent layer on top of rsync that accepts nearly all the command-line options available to rsync.
stream-lib is a Java library for summarizing data in streams for which it is infeasible to store all events. More specifically, there are classes for estimating: cardinality (i.e. counting things); set membership; top-k elements and frequency.
The high performance computing group at RENCI is developing an efficient hierarchical OpenMP implementation for the open-source Qthreads runtime library. Efficient scheduling of tasks on modern multi-socket many-core shared memory systems requires consideration of an increasingly complex memory hierarchy. In a traditional work-stealing framework, each core is assigned a double-ended queue (a dequeue). One side of the dequeue supports sequential operations while the other side supports concurrent modification. In contrast, our runtime assigns several cores to a single dequeue. For these cores on the same chip, the LIFO task scheduling allows exploitation of cache locality between sibling tasks as well as between a parent task and its newly created child tasks. My contribution is the design and implementation of an array-based lock-free dequeue that supports concurrent modification on both ends of the data structure.
The paper “OpenMP Task Scheduling Strategies for Multicore NUMA Systems” has been accepted for publication in the peer-reviewed journal International Jounral of High Performance Computing Applications.
A haplotype is a DNA sequence that has been inherited from one parent. Each person possesses two haplotypes for most regions of the genome. The process of deducing haplotype information based on genotype data is known as haplotype phasing. Haplotype information is important in untangling the heritability of traits of interest such as generic disorders. We can use genetic markers that are uncommon in the population as beacons to identify common ancestors for individuals that are not known to be related. The design of this phasing strategy is ongoing. My responsibilities involve scaling our phasing algorithm to thousands of individuals and hundreds of thousands of genetic markers per individual.
The OpenMx Project intends to rewrite and extend the popular statistical package Mx to address the challenges facing a large range of modern statistical problems such as: (i) the difficulty of measuring behavioral traits; (ii) the availability of technologies - such as such as magnetic resonance imaging, continuous physiological monitoring and microarrays - which generate extremely large amounts of data often with complex time-dependent patterning; (iii) increased sophistication in the statistical models used to analyze the data. To address these problems, the Mx Structural Equation Modeling software will be rewritten so as to: (i) split OpenMx into modules that interoperate with the R statistical package; (ii) release OpenMx as open source so as to provide a stable path for future maintenance and development; (iii) integrate OpenMx with the Swift (formerly VDL) parallel workflow software.
The paper “OpenMx: An Open Source Extended Structural Equation Modeling Framework” has been accepted for publication in the peer-reviewed journal Psychometrika, the print journal of the Psychometric Society.
The power wall, the ILP wall, and the memory wall are driving a trend from implicitly parallel architectures towards explicitly parallel architectures. The memory wall has been identified as one of the fundamental challenges to high-performance concurrent computing. The design of cache-conscious concurrent data structures for many-core systems will show significant performance improvements over the state of the art in concurrent data structure designs for those applications that must contend with the deleterious effects of the memory wall. The design of cache-conscious, linearizable concurrent data structures has advantageous properties that can be measured across multiple architecture platforms. My dissertation research fills the gap in cache-conscious concurrent data structures by providing concurrent algorithms that implement an ordered set abstract data type. The dense skip tree is a randomized data structure that has been designed to probabilistically exploit spatial locality of reference. The dense skip tree causes fewer cache misses than self-balancing binary search trees by probabilistically aggregating consecutive sequences of keys into contiguous regions of memory. The primary contributions of my thesis are the optimistic skip tree algorithm, the lock-free concurrent skip tree algorithm, and the lock-free concurrent HAT trie algorithm.
I successfully defended my dissertation on April 18, 2011. A copy of the thesis is available online.
Fortress was a programming language designed for high-performance computing with high programmability. Fortress will support features such as transactions, specification of locality, and implicit parallel computation as integral features built into the core of the language. Features such as the Fortress component system and test framework facilitate program assembly and testing, and enable powerful compiler optimizations across library boundaries. The syntax and type system of Fortress are custom-tailored to modern HPC programming, supporting mathematical notation and static checking of properties such as physical units and dimensions, static type checking of multidimensional arrays and matrices, and definitions of domain-specific language syntax in libraries.