Research in Progress

• Peer-to-Peer Parallel I/O on Grids

  Grids connect geographically distributed computing resources together to provide high performance computing power for users across the nation and the world. For data-intensive Grid applications, the traditional client-server storage model can severely limit system performance. To overcome this centralized I/O barrier, we are working on a peer-to-peer storage system on the Grids.

• Load Balancing Heterogeneous Storage Systems

  We are also developing a workload balancing algorithm for heterogeneous storage environments. Parallel computing systems uausally have heterogeneous storage devices:

  • The Grid is inherently a heterogeneous system;
  • Clusters usually have heterogeneous storage devices;
  • File-system aging problem introduces heterogeneity to the system.

Previous Projects

• I/O Workload Characterization

  Investigated static (SUIF compiler-based) and dynamic (profile-guided) approaches to analyze spatial and temporal parallel I/O access patterns. We concluded that parallel scientific applications always exhibit very regular and highly predicable I/O access patterns.

• Parallel I/O Modeling and Simulation

  We have developed an execution-driven parallel simulator for performance prediction of both communication and I/O intensive applications. This simulator is built on top of DiskSim and can accurately predict performance of parallel applications as a function of underlying storage architectures and network connections.

• Profile-guided I/O Partitioning

  To achieve scalable I/O performance of parallel applications, it is important to parallelize I/O streams at both the application and file levels. In this project, we designed a scalable high performance I/O system by partitioning data files across multiple storage devices, in an attempt to maximize I/O parallelism. For the applications we have studied, partitioned I/O reduced overall execution time by as much as 82% compared with MPI collective I/O.

• Parallel SDFMM

  We have parallelized a Steepest Descent Fast Multipole Method (SDFMM) subsurface sensing and imaging application within the CenSSIS project, and significantly reduced its runtime on a 32-processor Beowulf cluster.

• I/O Tracing

  In past joint research between EMC Corporation, I-Tech and NUCAR we investigated new tracing tools for the I/O domain. These tools allow unlimited length traces of SCSI bus activity to be captured in real time. We are working on a new tracing tool able to capture multiple streams of I/O concurrently, and have used this tool to capture several large IO traces of synthetic, as well as real, workloads. We are using these traces to investigate new architectural features in the I/O domain. This project will also study the characteristics of a number of datamining applications, hosted in an Oracle database.