Computer Architecture and Performance Modeling (CAP)

Performance analysis of computer system architectures exploiting new technology trends using advanced workload characterization and modeling techniques.

Project Details

Computer Architecture and Performance Modeling (CAP)

Computer Architecture and Performance Modeling (CAP)

Performance analysis of computer system architectures exploiting new technology trends using advanced workload characterization and modeling techniques.

Project Overview


Modeling and Analysis

The CAP group does full-system performance modeling and analysis, using analytical techniques such as queuing models and bounds analysis as well as simulation. We have built and used detailed, cycle-accurate models as well as high-level abstract models to understand the performance effects of architectural design decisions. Our group has worked with systems and applications in many areas of interest to Oracle customers, including enterprise software, databases, and scientific computing / High-Performance Computing (HPC) applications. We are currently working to improve OCI compute performance.

Workload Characterization

The CAP group has significant experience with application profiling, tracing, and analysis tools. This includes the development of custom tools and methodologies to fit project or workload-specific modeling goals. We have developed and extended tools that have been adopted by other groups within Oracle, and we contribute improvements to open-source analysis tools.

Recent Projects

The CAP group has performed modeling and characterization as part of the following recent projects:

  • Bringing world-class dense-linear algebra (DLA) computational capability to Oracle: through our collaboration with the UT Austin Science of High Performance Computing Group, we have optimized BLIS (BLAS-like Instantiation Software, the world’s highest performance DLA library) for all of Oracle’s compute platforms. DLA is very important for the performance of machine learning and HPC applications.
  • OCI node analysis, characterization, and performance improvement: using BLIS as a representative of a cooperative multithreading application like the Oracle DB, we have analyzed the performance of OCI nodes and provided feedback to OCI, the BLIS community, and to vendors. Ampere, Oracle’s vendor for our ARM nodes, has made changes to their microprocessors and microprocessor roadmap base on our analysis.
  • Java-BLIS: by integrating BLIS with Oracle JVMs, we have demonstrated that Java is a first-class platform for machine learning and provided feedback for/use of Java’s new Foreign Function and Memory interface.
  • Cross-Language Microbenchmark Harness (CLAMH): we created and released an open source test harness that allows GraalVM users/prospects to create and run performance benchmarks that demonstrate the excellent performance of GraalVM.