.pdf
Kokkos Tools: Profiling Tuning Debugging
1. Introduction
Kokkos Tools represent a sophisticated suite of utilities designed to enhance the development and optimization of high-performance computing applications. These tools leverage Kokkos' built-in instrumentation to provide developers with powerful capabilities for profiling, debugging, and tuning their code across diverse hardware architectures.
2. Kokkos Tools and Built-in Instrumentation
The Need for Kokkos-aware Tools :
-
Modern heterogeneous computing environments present complex challenges for performance analysis and optimization.
-
Traditional profiling and debugging tools often lack context-specific information for Kokkos applications.
-
Kokkos-aware tools bridge this gap by interfacing directly with the Kokkos runtime, providing more meaningful insights.
How Instrumentation Helps ? :
-
Kokkos' built-in instrumentation allows for non-intrusive gathering of detailed execution information.
-
It tracks critical events such as kernel launches and memory operations without requiring source code modifications.
-
This approach minimizes impact on application behavior while still offering rich performance data.
Simple Profiling Tools :
-
KernelLogger: Helps developers localize errors and verify runtime flow by printing Kokkos operations as they occur [1].
-
SimpleKernelTimer: Measures time spent in kernels, identifying hotspots and aiding in performance optimization [1].
-
MemoryEvents: Tracks memory-related events, helping identify issues like excessive temporary allocations [1].
Simple Debugging Tools :
-
KernelLogger: Acts as a debugging tool by inserting fences that check for errors and printing Kokkos operations [4].
-
These tools can help pinpoint issues in kernel execution and memory management, crucial for complex parallel applications.
3. Vendor and Independent Profiling GUIs
What Connectors Provide ? :
-
Connectors translate Kokkos instrumentation for use with vendor-specific and independent profiling tools.
-
They bridge the gap between Kokkos' internal instrumentation and external profiling interfaces.
-
This allows developers to use familiar tools while gaining Kokkos-specific insights.
Available Tools :
-
nvtx-connector: Interfaces with NVIDIA tools like Nsight Compute, translating KokkosP hooks into NVTX instrumentation [4].
-
vtune-focused-connector: Enables integration with Intel’s VTune profiler for detailed performance analysis on Intel architectures.
-
TAU (Tuning and Analysis Utilities): Offers built-in support for Kokkos without requiring a separate connector [2].
4. Tuning
As applications grow in complexity, the need for tuning becomes increasingly apparent. Kokkos recognizes this need and provides autotuning hooks to help developers optimize their code for different architectures and workloads.
The necessity for tuning is evident when considering the myriad of parameters that can affect performance. For instance, in a sparse matrix-vector multiplication (SpMV) implementation, factors such as the number of rows per team, team size, and vector length can significantly impact performance across different hardware [5]. Manually determining optimal values for these parameters across various architectures is a daunting and time-consuming task.
5. Custom Tools
The KokkosP Hooks :
-
KokkosP interface exposes hooks corresponding to various Kokkos runtime events.
-
These hooks include kernel launches, memory operations, and region entries/exits.
Callback Registration Inside the Application :
-
Developers implement callback functions for relevant KokkosP hooks.
-
These callbacks are registered with the Kokkos runtime to be invoked at appropriate execution points.
Throwaway Debugging Tools :
-
Lightweight, purpose-built tools can be quickly implemented for specific debugging scenarios.
-
Example: A tool to log memory allocations exceeding a certain size to identify potential memory leaks.