IBM Almaden Research Center

There is an increasing trend for IT configurations to include very powerful storage servers. At the same time, the traditional shared-nothing style of parallel query processing has serious scalability limitations (see VLDB'05 paper below). So, we are investigating an alternate componentization of the database system archtitecture where some data intensive operations are embedded into processors close to the disk: SAN controllers, RAID controllers, or even lower, and other operations are pulled out into separate compute blades.

We are investigating Data In The Network (DITN), a new method of parallel querying based on dynamic outsourcing of join processing tasks to non-dedicated, heterogeneous computers. In DITN, partitioning is not the means of parallelism. Data layout decisions are taken outside the scope of the DBMS, and handled within the storage software; query processors see a Data-In-The-Network image.

This allows gradual scaleout as the workload grows, by using non-dedicated computers. A typical operator in a parallel query plan is the Exchange operator that migrates tuples between nodes. One of our main arguments is that Exchange is unsuitable for non-dedicated machines because it poorly addresses node heterogeneity, and is vulnerable to failures or load spikes during query execution. DITN uses an alternate intra-fragment parallelism where each node executes an independent select-project-join- aggregate-group by block, with no tuple exchange between nodes. This method cleanly handles heterogeneous nodes, and well adapts during execution to node failures or load spikes. Initial experiments suggest that DITN performs competitively with a traditional configuration of dedicated machines and well-partitioned data for up to 10 processors at least. At the same time, DITN gives significant flexibility in terms of gradual scaleout and handling of heterogeneity, load bursts, and failures.

Blue Gene/L, IBM's high-performance multiprocessor computer, is traditionally used within scientific computing for the physical sciences such as chemistry and biology. Our goal in this project is to design, implement, and evaluate middleware components that will allow financial applications to be run on Blue Gene/L. This work was motivated by the requirements of a specific proprietary application involving an asset portfolio Value-at-Risk (VaR) computation using Monte Carlo simulation.

Financial applications present technical and usage requirements beyond that of scientific computing, including dependence on external data sources such as SQL databases, interaction within larger business workflows, and separation of high-level service specifications from low-level service provisioning. In order to support such applications on Blue Gene/L, our middleware, written in Java and running on a Blue Gene front-end node, provides a number of core features such as an automated data extraction and staging gateway, a standardized high-level job specification schema, a well-defined web services (SOAP) API for interoperability with other applications, and a secure HTML/JSP web-based interface suitable for general users (i.e., non-developers).

We designed and implemented a novel architecture for data processing based on a functional fusion between a data layer and a computation layer. We show how such an architecture can be leveraged to offer significant speedups for data processing jobs such as data analysis and mining over large data sets.

One novel contribution of our solution is its data-driven approach. The computation infrastructure is controlled from within the data layer. Grid computational job submission events are based within the query processor on the DBMS side and in effect controlled by the data processing job to be performed. This approach allows the early deployment of on-the-fly data aggregation techniques, minimizing the amount of data to be transfered to/from compute nodes and is in stark contrast to existing Grid solutions that interact with data layers mainly as external "storage".

• Evolving Toward the Perfect Schedule: Co-scheduling Job Assignments and Data Replication in Wide-Area Systems Using a Genetic Algorithm. (T. Phan, K. Ranganathan, and R. Sion) Proceedings of the Workshop on Job Scheduling Strategies for Parallel Processing (JSSPP), 2005.
• Parallel Querying with Non-Dedicated Computers.  (V. Raman, W. Han, and I. Narang). International Conference on Very Large Data Bases (VLDB), 2005.
• XG: A Data-driven Computation Grid for Enterprise-Scale Mining. (R. Sion, R. Natarajan, I. Narang, W-S. Li, and T. Phan) Proceedings of DEXA, 2005.
• Spreading the Load Using Consistent Hashing: A Preliminary Report. (Garret Swart) ISPDC/HeteroPar 2004