Using Kyvos resources with GPU-acceleration

Kyvos introduces GPU-accelerated querying to deliver lightning-fast performance for large-scale analytical workloads. This capability is designed to work seamlessly across all raw data sources supported by Kyvos' No-Spark architecture, making it ideal for high-performance, modern BI environments.

How It Works

To enable GPU-based query optimization, Kyvos relies on a key configuration property:

• Property: kyvos.parquet.cuboid.enable

• Value: true

When this property is set to true, Kyvos stores aggregated cuboids (pre-aggregated data structures generated during semantic model builds) in Parquet format—a highly efficient, columnar storage format optimized for GPU processing.

Key Benefits

• Accelerated Query Performance: GPU-based execution significantly reduces query response times.

• Optimized Storage: Parquet format enables efficient disk usage and faster data access.

• Broad Compatibility: Works across all raw data sources within the No-Spark architecture.

• Scalability: Designed to handle massive datasets and complex analytical models.

Deployment Requirements

• Supported Platform: GPU acceleration is currently supported in AWS Enterprise environments only.

• Instance Type: Kyvos recommends using the g6.2xlarge instance type when creating Kyvos resources to enable GPU-based query execution.

Upgrade and Patch Guidelines

For AWS, during upgrades and Patch Kyvos and its components, ensure the drutdb_jdbc file is uploaded using either the Path-based or File-based method via the Upgrades and Patch page in Kyvos Manager. This enables detection and visibility of the hardware and execution configurations required for GPU support.

Points to know

• From Kyvos 2025.9 onwards, GPU enabled querying now supports:

  • For SQL and MDX function: Join and Limit

  • For SQL function: Coalesce

• GPU enabled queries also:

  • Support for Expressions on Date/Time Data Types: Enables the use of expressions to extract components such as month, year, or day from date/time values, including functions like strptime(date_column), year(date_column), month(date_column), and day(date_column).

  • GPU Utilization Filter in Kyvos Monitor: Introduces a “GPU Utilization” filter in the Completed Process tab of Kyvos Monitor, allowing users to trace and monitor Kyvos queries based on GPU usage percentage.

  • Support for the WITH Clause in SQL: Adds support for using the WITH clause within SELECT statements. However:

    • Multiple WITH clauses at the same hierarchical level are not supported.

    • Reusing a single WITH clause multiple times in an outer query is not supported.

  • Support for Result Set-Based Querying: Allows queries to operate on intermediate result sets.

  • Support for ORDER BY in SQL Queries: Enables sorting of results using the ORDER BY clause. If data volume exceeds available GPU resources, query execution will automatically fall back to CPU processing.

  • Support for the CAST Function in SQL Queries: Adds support for the SQL CAST function for data type conversion within queries.

• If GPU execution is unavailable, queries will automatically fall back to CPU execution without failure.

• The Query Details dialog box displays the execution mode—CPU or Hybrid—along with the percentage of the query processed by GPU.

• Kyvos currently supports only NVIDIA GPUs. Other GPU types are not supported.

• Columns with the BigDecimal data type are not supported for GPU-based querying.

• Only Base Aggregates are supported. Smart Aggregates are not supported.

• GPU-based querying is supported only for no-Spark semantic models. Queries on Spark-based models are not supported.

• Kyvos will utilize only 80% of the available GPU memory to maintain a safe memory threshold and ensure that sufficient GPU resources remain available for other processes.

• After a fresh deployment with GPU support enabled, the query engine may require an additional restart to begin utilizing GPU resources effectively.

Known Behavior

• If a query cannot be scheduled on the GPU within 2 seconds due to high demand, it will be removed from the GPU queue and executed on the CPU instead.

• GPU-based querying is supported only on the amd64 (x86_64) architecture.
  arm64 architectures are not supported.

• When semantic models are built with Parquet-based cuboids, the following features are not supported:

  • Advanced hierarchies

  • Many-to-many relationships

  • Slowly Changing Dimensions (SCD)

  • Semi-additive measures

Read more advanced configurations:

• AS_GPU_MIN_CHUNK_SIZE 

• AS_GPU_MAX_CHUNK_SIZE