SMA CS Research

Supporting Task-Dependencies on the ALiCE Grid

Student : Asankha Chamath Perera
SMA Supervisors : Assoc Prof Teo Yong Meng (Singapore) & Prof Alan Edelman (MIT)

Abstract:

This dissertation investigates supporting parallel Java applications with task-dependencies in the distributed shared memory ALiCE Grid. Programming in ALiCE is supported by a set of programming templates, which hide the complexities of the Grid infrastructure. The main templates available within the ALiCE system support the definition of computationally intensive and parallelizable Tasks; and Task Generators which controls the application flow. The Task Generators and Tasks of an application executes within the different Java virtual machines of the Producer nodes that volunteer computational resources to the ALiCE Grid. A new set of programming templates allows programmers to develop applications with task-dependencies, with elegant and efficient code, while still maintaining backward compatibility to support existing applications. The new enhancements to the templates, which are modeled similarly to Java threads, allows programmers who are new to Grid computing to easily harness the powerful capabilities presented, with minimal and simple code.

Support has been implemented to allow tasks to spawn sub tasks, and wait for their completion, synchronize at barriers, and optionally collect their results and detect exceptions encountered during remote execution. The enhanced runtime system supports the concurrent execution of multiple tasks and task generator instances at Producer nodes, allowing blocking jobs to remain active until they are resumed, with minimum overhead to the system. The new templates allow the submission of the same application in batch and interactive modes of execution without code duplication or recompilation. Due to on-line output file replication support implemented for interactive mode job submissions, the implementation of the visualization component has been separated from the Grid application.

SVM Active Learning and Applications in Image Retrieval Systems

Student : Deng Kun
SMA Supervisors : Assoc Prof Lee Wee Sun (Singapore) & Prof Leslie Pack Kaelbling (MIT)

Abstract:

An SVM active learning based image retrieval system is implemented for research purpose. Some of the interesting properties of the system are studied and experiments show active learning method outperforms normal SVM learning and many other relevance-feedback algorithms if the query concept is “separable”. Some algorithms with different SVM sampling criteria are tried out. They are threshold based random sampling and most-orthogonality based sampling methods. However experiments also show engineering effective active learning querying components is still a subtle task as many seemingly reasonable sampling criteria for choosing a pool instance can even perform worse than random sampling.

Reinforcement Learning in Optimal Ventilator Control

Student : Ding Yang
SMA Supervisors : Assoc Prof Leong Tze Yun (Singapore) & Prof Leslie Pack Kaelbling (MIT)

Abstract:

This dissertation reports using reinforcement learning and other artificial intelligence techniques to get an optimal policy in controlling ventilators. It provides a general method to help decision-makings during ventilation in hospitals. Ventilator control is a complex and challenging problem which involves lots of attributes, both patient physiological features and mechanical settings. My work is focused on Assistant Control ventilation mode, which is simpler and most widely used in clinical practice. A new system structure to cope with this problem is described, with a physiological model and a policy learner as two major parts in my work. Instead of developing the respiratory system functions, my work is based on the clinical data from Intensive Care Units in a local hospital. Necessary processing to the raw data is suggested, which is decided by medical domain knowledge. I use a simple statistical model to estimate human respiratory model. Interestingly, this physiological model fits the testing data well. However, another approach using artificial neural networks gets even better results. I adopt the neural network model in the second part of my work. Q-learning is used to get an optimal policy, which aims to control the ventilator setting. The ventilator control problem is converted to a model-based problem in reinforcement learning. Value iterations are used to determine the optimal control policy. Experiments have been done to show the feasibility of this approach. Due to some limitations such as expensive computation, this method doesn't extend effectively to large state space. The advantages and disadvantages with possible improvements are also discussed.

Distributed Matlab on the ALICE Grid

Student : Lee Yih
SMA Supervisors : Assoc Prof Teo Yong Meng (Singapore) & Prof Alan Edelman (MIT)

Abstract:

Current technology trends show an increasing interest in parallel computing and distributed computing. In particular, grid computing is widely believed to be one of the emerging technologies that will change the world. The ALiCE grid technology being developed here at NUS is based on distributed shared memory (DSM) architecture. While there are currently at least 27 parallel Matlab implementations, there are so far no Matlab implementations on a grid system. This thesis describes the design and implementation of Matlab*G, a distributed Matlab on the ALiCE grid. The design draws on the idea of Matlab*P, a parallel Matlab developed at MIT. This thesis also describes Matlab*J, a distributed Matlab on DSM. Although Matlab*J is a component of Matlab*G, it functions independently from ALiCE and can be used on its own without ALiCE. Matlab*J is a first distributed Matlab on a DSM system. Although the implementation of Matlab*J is on Sun JavaSpaces, the design can be modified to use other DSM systems such as GigaSpaces. The ALiCE system used in this thesis is also based on Sun JavaSpaces but it also supports other DSM systems.

Component Based Query Processor

Student : Poorna Subhash Pullipudi
SMA Supervisors : Assoc Prof Tan Kian Lee (Singapore) & Prof Stuart Madnick (MIT)

Abstract:

The present day commercial database systems are evolved with time by adding many features and functionality. Therefore, these database systems are very large in size and monolithic in structure. In this project we considered component based approach for developing database systems. We centered our focus on the query processor subsystem of the database system. We proposed a component based layered architecture for query processor with the granularity of the components at operator level. We also proposed guidelines in developing the components and composing the database system from off the shelf components. We evaluated the ease of the approach by implementing few components and composing the system. the component based approach for query processor guarantees minimum footprint, predictability and extendibility.

Sized Regions for Real-Time Java

Student : Sin Mong Leng
SMA Supervisors : Assoc Prof Chin Wei Ngan (Singapore) & Assoc Prof Martin Rinard (MIT)

Abstract:

This report introduces Sized Regions for Real-Time Java, which is a big bring together of sized typed analysis, regions and Java. Annotated Sized Regions is used to check the correctness of regions usage and detect potential dangling pointers due to bad programming. Sized regions use sized types to check annotated size information thus paving the way for size inferencing at a later stage. In this report we identify a subset of Java for our analysis. Annotated region syntax is checked for correctness. Sized annotations are introduced for more checking and gathering of size information. Various issues like fixed-point analysis and structural aliasing are discussed briefly in this report.

Dynamic Optimization via Hyper-Threading

Student : Varun Talwar
SMA Supervisors : Assoc Prof Wong Weng Fai (Singapore) & Prof Saman Amarasinghe (MIT)

Abstract:

Dynamic linking, translation and compilation have enabled software to dynamically adapt to new environments. This is becoming increasingly important with the distribution of software via the internet. Dynamic optimization is an attempt to optimize code as it is run. This does away with the need to access the full source code of the application and also enable optimizations across dynamically linked code. Since dynamic optimization is done at runtime, the main challenge in engineering such a system is to keep the overhead involved minimal so that it doesn’t outweigh any benefits that may be gained from the optimizations. This thesis introduces the concept of sideline optimization whereby a low priority thread is spawned to perform code optimization while the main application runs under minimal interference from the dynamic optimization runtime system. A multithreaded extension to the dynamic optimization framework DynamoRIO is proposed. The performance of this system on a traditional symmetric multiprocessor system as well as on a simultaneous multithreading architecture, as realized by the Intel hyperthreading processors, is studied. The experiments show that for some applications, a speedup of up to 36% over the original statically optimized code is achievable.

Intelligent Information Integration for Colorectal Cancer Management

Student : Wei Songjie
SMA Supervisors : Assoc Prof Leong Tze Yun (Singapore) & Prof Stuart Madnick (MIT)

Abstract:

A significant proportion of the information required for biomedical research is recorded in the on-line citation databases such as MEDLINE. Such information is important for clinical decision analysis. Here we present an automatic approach to extract semantic relation from different data sources to facilitate decision model construction for colorectal cancer management. We make use of the Unified Medical Language System (UMLS) for both overcoming concept heterogeneity and constructing candidate semantic relations. Each relation is evaluated by sending query back to the data source and getting the statistical evidence about this relation. The abstracts of citations are also parsed to find supporting evidence. Decision elements and possible relationships among them may be derived from the semantic relations to support construction of a clinical decision model. The final relation diagram can also help to construct multilevel influence diagram, which is useful for decision-driven model construction for clinical problems.