Hopkins Storage Systems Lab

Currently, resource sharing and performance management for high-performance distributed file systems pursue incomplete or incorrect goals. Server-centric performance metrics, such as I/O latency and throughput, are insensitive to client state and cannot distinguish the urgency of file system operations by I/O type or among different clients. Storage networks rely on flow control to apportion bandwidth, which is irrelevant for all configurations that are not network bound. Client acts selfishly, trying to maximize their throughput, which congests server and network resources. 

In response, we are developing holistic resource management algorithms for high-performance distributed file systems that use online auctions to maximize application-perceived performance. The system is holistic in that it manages all resources, including network bandwidth, server I/O (throughput and IOPS), server CPU, and client and server memory utilization. Online auctions unify multiple heterogeneous resources in a single pricing model, which allows the system to adapt to different configurations and time-varying or workload-varying resource constraints. The focus on application-perceived performance ensures that optimization goals benefit the system’s users (not its servers). 

We have also introduced a new dimension in resource management for distributed file systems by managing adaptively the global allocation of memory among clients and servers and the assignment of memory to read caching and write buffering. Emerging high-speed networking technologies bring memory limitations to the forefront, exposing systems to throughput crashes and applications stalls. 

We are implementing these techniques in the congestion-aware network file system (CA-NFS), which extends the NFSv4 file system to implement auctions and pricing and the Linux memory manager to implement adaptive read/write scheduling and memory management. 

The Sciences are encountering a data avalanche problem in which improvements in physical instruments and data pipelines lead to an exponential growth in data size. Federated and clustered databases are attractive solutions for exploring the resulting massive, widely-distributed data. One example is the SkyQuery federation of Astronomy databases. To ensure high job throughput and prevent starvation of traditional workloads, we propose LifeRaft, a query processing disciplines to eliminate redundant I/O for scan-based queries. Sites in SkyQuery service millions of queries each month in which queries may execute for several hours or an entire day. We reduce redundant I/O requests to the disk from concurrent queries by relaxing in-order scheduling. Specifically, rather than execute queries in arrival order, we interleave I/O requests from multiple queries based on contention for shared data. While our approach increases response time of certain queries, we observe over two fold improvement in system throughput for Astronomy workloads.

Data size and geography in SkyQuery dictate that transmitting data takes a long time and has a profound impact on query performance. Each site may produce hundreds of megabytes of data that are sent to other sites to be joined before results are delivered to the scientists. We devised algorithms that incorporate network structure, such as the throughput of paths and clusters of sites, and account for data access requirements in query scheduling. By exploiting excess capacity in the network and avoiding transfers across large geographies, we achieve orders of magnitude benefit for queries that join ten or more databases.

Finally, we devise automated physical schema design tools for the management of large scale scientific databases. Many current tools are offline and demand that DBAs identify a representative workload for tuning, decide when tuning is required, and guesstimate the relative benefits of changing the schema design. We are currently developing AdaptPD, a tool that continuously monitors the workload and adapts the schema design to suit the incoming workload. It models schema design as a metrical task system and also includes query and transition cost estimation modules to ensure that the tuning process is light weight.

Placeholder.

This CBEO (Chesapeake Bay Environmental Observatory) integrates more than 15 model-generated and observational data sets from the Chesapeake Bay. It allows scientists to explore, compare, and correlate data from buoys, cruises, over-flight, satellites, shallow-water monitoring, and high-resolution hydrodynamic and water-quality models. The Chesapeake is the most intensively studied estuary in the United States.  Providing scientists with the ability to
compare, correlate, and join data among these data sets has transformed our understanding of critical Bay processes, such as hypoxia or the de-oxygenated dead zone, loss of coastal grasses, and how sewage runoff
impacts geochemistry.

In the HSSL, we are building the databases that store and manage heterogeneous spatial data sets of the CBEO and provide access paths and data organization techniques specifically suited to the Chesapeake.

We are constructing a Community Databases for the storage and analysis in database clusters of multi-scale data sets produced on high-performance computing platforms. Using the databases, scientists can perform feature extraction, data mining, and visualization. We term this approach to computational science remote immersive analysis in that scientists have access to the entire computational output interactively over the Internet. The databases are available for public use through a Web services interface that connects the data to the computational tools scientists use most, such as Matlab, C and Fortran.

At present, we have 30 TB of forced isotropic turbulence data computed by direct numerical simulation online.  Our future plans include channel-boundary flows and magneto-hydrodynamic turbulence.  Please visit the project Web page and access the database at http://turbulence.pha.jhu.edu/.