High Performance Computing
The High Performance Computing (HPC) group at IBM Research - India is engaged in designing and analyzing cutting edge parallel programs and improving the performance of engineering, scientific, and business applications on high performance platforms such as the IBM Blue Gene supercomputer. The group has following major focus areas:
Following sections provide more details on the research projects in the HPC area.
HPC Challenge Benchmarks
The HPC Challenge (HPCC), a suite of seven benchmarks, is fast evolving as a standard for evaluating the performance of supercomputers across a spectrum of real-world applications. The HPC team at IBM Research - India is actively involved in performance optimization and tuning of the HPC Challenge benchmarks on high-end systems, such as the Blue Gene supercomputers and POWER 7 based systems. The optimizations include designing new algorithms, data structures and other intricate techniques for distributed memory and multicore architectures. This group has optimized the STREAM, RandomAccess, Fast Fourier Transform and Transpose benchmarks for various systems. The HPC Challenge RandomAccess benchmark optimized by this team has won the HPCC Class I award at Supercomputing for all the 6 years (2005-2010).
The HPC team is also involved in optimizing various scientific applications that require high-end systems for large scale processing. These applications include molecular dynamics simulation packages such as VASP, weather modeling packages, smart grid applications such as contingency analysis, etc. on the Blue Gene supercomputers.
Smarter Wireless Solutions
Smarter Wireless Solutions group focuses its activities on architectures and algorithms for 3G and 4G telecommunication networks and private enterprise networks. The group is co-leading research on the Wireless-IT convergence which is emerging as an area of great opportunity for IBM. Wireless-IT convergence explores the opportunities for improving the performance and impact of future wireless networks through intelligent use of Information Technology (IT).
The main research threads in the group are:
Wireless/IT Solutions for Wireless Networks:
Increasing number of mobile devices and growing multimedia traffic on wireless networks introduces severe bottleneck at the wireless link, backhaul and the core network of the underlying infrastructure. The group is working on cost-effective solutions to reduce this bottleneck using advanced wireless and IT optimizations. Other research areas include improving quality-of-service for multimedia streaming over broadband wireless networks, and optimizations for machine-to-machine communications.
Researchers: Ashok Ambati, Umamaheshwari Devi, Partha Dutta, Kunal Korgaonkar, Ramana Polavarapu, Malolan Chetlur and Shivkumar Kalyanaraman.
Wireless Network Cloud:
The Wireless Network Cloud (WNC) is a novel wireless system architecture that leverages developments in cloud computing and emerging wireless technologies, such as Software Radio and Remote Radio Head technology. WNC allows separation of hardware and software development for different wireless standards and enables cost savings through infrastructure sharing between operators. It also opens up many new business models for network access and service providers. This project is in collaboration with IBM Research - China to define the WNC system requirements and architecture and build a WNC proof-of-concept.
Researchers: Parul Gupta, Mukundan Madhavan, Dheeraj Sreedhar, Malolan Chetlur and Shivkumar Kalyanaraman, along with IBM Research - China
Networking requirements for datacenters are changing with cloud computing. As cloud datacenters scale to tens of thousands of servers, the traditional single rooted, hierarchical and over-provisioned datacenter network starts becoming a bottleneck from a performance and manageability perspective. The cloud networking effort at IBM Research - India is exploring new datacenter network architectures specifically aimed for cloud computing. One of the problems that we are currently exploring is cloud traffic modeling and simulation. The problem is inherently complex due to the numerous layers interacting in the cloud and high volatility in network traffic. Most of the recent datacenter architectures have been evaluated on modest real testbeds (upto 100 servers with 10-15 switches). It is hard to predict performance at 1000s of servers with 100s of switches. Another important aspect of our work is exploring novel methods of doing networking functions in software.
Researchers: Vijay Mann, Partha Dutta and Shivkumar Kalyanaraman, along with IBM Research-China
Smarter Energy Solution
Smarter Energy Research group is focused on developing technologies for power grids that can supply electricity to everyone with minimal impact on the environment. In particular, we are working on the following aspects:
We collaborate with several industrial organizations (Sampol and Tecnalia among others) and academic partners including Athens University of Economics and Business, CMU, EPFL, IIIT Delhi, IIT Madras, IIT Kharagpur, Lulea University, Monash University, RWTH, Universiti Brunei Darussalam (UBD) and University of Waterloo.
Researchers: Deva P. Seetharam, Chumki Basu, Jagabondhu Hazra, Kejitan Dontas, Vijay Arya, Sunil K Ghai, Kaushik Das, Tanuja Ganu, Balakrishnan Narayanaswamy, Ashok P Prakash, Swarnalatha Mylavarapu, Zainul M Charbiwala, Mohit Jain.
Example research threads in the group are:
Demand Response [DR] programs along with the emergence of technology to support DR are known to be effective in reducing electricity consumption. However, the process of designing effective DR programs is not well understood. Research related to demand response and demand side management has addressed a suite of issues including differential pricing, incentive schemes, consumption profiling, and eco-feedback, etc. Despite these efforts, there is a lack of understanding of factors (such as weather, local events and consumer context) that influence electricity consumption and demand side participation. Understanding these factors is crucial to success of a DR program. The aim of this work is to address this gap by developing a monitoring, inference and observation framework to understand fine-grained consumption patterns, to establish relationships between users' consumption patterns and their contexts and to evaluate cost-effective pricing/incentive schemes. The insights gained will be presented to DR designers using a visual analytics framework that can help the designers and policy makers choose optimal program features such as pricing models, incentives, user interfaces, etc.
Stream Computing Based Synchrophasor Application For Power Grids
This work explores the application of stream computing analytics framework to high speed synchrophasor data for real time monitoring and control of electric grid. High volume streaming synchrophasor data from geographically distributed grid sensors (namely, Phasor Measurement Units) are collected, synchronized, aggregated when required and analyzed using a stream computing platform to estimate the grid stability in real time. This real time stability monitoring scheme helps the grid operators to take preventive or corrective measures ahead of time to mitigate any disturbance before they develop into wide-spread.
High Performance Analytics
The High Performance Analytics group at IBM Research - India is engaged in the design, implementation, performance modeling and analysis of the state-of-the-art parallel analytics algorithms and kernels. These analytics kernels are useful in Smarter Planet Domains such as Smart Telecom, Smart Grid/Security as well as scientific domains such as computational biology, computational astronomy and so forth. The underlying algorithms are optimized for multi-core and distributed architectures. This group is also engaged in distributed scheduling and related performance optimizations of runtime systems such as Map Reduce for multi-core and cluster architectures and Blue Gene L/P.
Following sections provide more details on the research projects in this group.
Optimized Map Reduce Runtime System
This project aims to design and implement optimizations for Map Reduce runtime system on next generation architectures such as multi-core clusters and supercomputers. The Map Reduce paradigm has become very popular for parallel execution of scientific and data mining applications. Performance optimizations for Map Reduce systems on multi-core architectures will enable efficient map reduce solutions in terms of cost, power and area.
Researchers: Ankur Narang, Jyothish Soman, Vikas K. Garg
Analytics Kernels for Multi-Core Architectures
Optimization of analytics kernels on next generation multi-core architectures has become an important area of research. As part of this project we are looking at analytics kernels that are relevant for domains such as Smart Telecom/Wireless, Computational Biology, Cryptographic applications and so forth. We aim to design optimized parallel algorithms for graph and data mining/machine learning algorithms for multi-core clusters, large SMPs, hybrid systems including GPUs and Power EN as well as supercomputers such as Blue Gene/P.
Researchers: Ankur Narang, Raj Gupta, Souvik Bhattacherjee, Abhinav Srivastava, Sheetal Lahabar
Highly Scalable Distributed Real-time Text Indexing & Search
Data Intensive Supercomputing requires processing massive amounts of data using high number of processors and possibly delivering real-time processing rates. This project studies the Text Indexing and Search algorithms on supercomputers such as IBM Blue Gene/L. It involves design of distributed data-structures and algorithms for massive scale text indexing and search. We have demonstrated indexing speed of 10GB/s using 4K processors of Blue Gene/L.
Researchers: Ankur Narang
Distributed Scheduling Algorithms for Parallel Computations
Scheduling dynamically unfolding computation graphs on distributed multi-core architectures is an important and challenging problem. This distributed scheduling algorithm needs to ensure deadlock free execution, should follow locality annotations and also optimize for low space, time and message complexity. PGAS languages such as X10 expect the programmer to specify locality annotations to obtain better performance. We study the design of distributed scheduling algorithms that follow locality while ensuring deadlock free execution under bounded space per place. This project involves both theoretical analysis of space, time and message bounds of scheduling algorithms and also implementation of the algorithms and studying their performance in terms of space-time trade-offs on multi-core cluster architectures.
Researchers: Ankur Narang. Abhinav Srivastava, Naga Praveen K. Katta.