Research Projects

Udayanga Wickramasinghe 


PhD Student
Department of Computer Science
School of Informatics and Computing
Indiana University
( GITHUB IU KEY KEY CV )


My long term goal is to explore and exploit different system/platform/computing architectures, to develop runtime and operating system techniques that enable low latency and high throughput communication pathways between critical system components. I believe this could greatly benefit present and future applications in High Performance Computing (HPC).

My current research interests revolve around optimizing group communication primitives for HPC to gain a better performance output. Irregularities at application run time are a major concern -- the effects of tiny fluctuations in different modules (Hardware, OS, Network, Libraries, applications) and synchronization, can cause cascading delays in both communication and computation phases of an application. I believe in exploring specialized runtimes and operating systems, and other avenues (modeling performance and reducing overheads, latency hiding, novel algorithms) to counter these undesired effects.

Research Projects

rmalloc
Udayanga Wickramasinghe, Andrew Lumsdaine

rmalloc is a high performance Dynamic RDMA memory allocator. rmalloc also has a thin RDMA access layer called ripe that expose RDMA network (i.e. read/write) operations.

Mobydick
Watshala Vithanage, Udayanga Wickramasinghe

MobyDick is a Multi-ExoKernel operating system named after the whale in Herman Melville's book "Moby-Dick; or, The Whale". MobyDick thinks of many core systems as distributed systems and worships End - to - End philosophy.

LibPhotonNBC - Collectives for Photon RDMA library
Udayanga Wickramasinghe, Ezra Kissel, Andrew Lumsdaine

Photon is a high performance portable (libibverbs, Cray-ugini, libfabric, Inifinipath-psm, etc) RDMA aware middleware library developed at Indiana University for Infiniband based communication in HPC clusters. LibPhotonNBC aims to implement collective routines on top of Photon utilizing its Put with completion events (PWC) abstraction.

HPX-5
HPX-5 is a (Asynchronous many task runtime) reference implementation of ParelleX specification. It is an open source, portable, performance-oriented runtime developed at CREST (Indiana University). HPX-5 provides a distributed programming model allowing programs to run unmodified on systems from a single SMP to large clusters and supercomputers.

Xinu for Zynq/Parallella Platform
Udayanga Wickramasinghe, Martin Swany, Andrew Lumsdaine

Porting Xinu (a light weight kernel developed at Purdue) to upcoming embedded platform Parallella baseboard (with Xilinx Zynq PS/ PL)

Flow
Udayanga Wickramasinghe, Greg Bronevetsky

‘Flow’ is a prototype implementation of a continuous or streaming query engine for key/value based data. Main purpose is to perform highly efficient distributed querying and processing of large graphs, etc data structures for HPC applications and hence achieve scalable performance over a significant number of of nodes.

Sight
Greg Bronevetsky, Udayanga Wickramasinghe

‘Sight’ is a log analysis tool developed at LLNL which helps analyze scientific and domain specific applications by instrumenting them to produce logs that can infer useful information to end user in terms of tracing, formatting/structuring, graphs, performance data (ie:- papi,etc), etc. Features inlcude stream based logging for sight by plugging in communication infrastructure such as MRNet.

HybridMPI
Andrew Friedley, Greg Bronevetsky, Udayanga Wickramasinghe, Andrew Lumsdaine, Torsten Hoefler

Hybrid MPI (HMPI) is a library containing research on optimizing intra-node message passing communication for modern multi-core systems.