Intel® VTune™ Profiler

User Guide

ID 766319
Date 10/31/2024
Public
Document Table of Contents

knob

Set configuration options for the specified analysis type or collector type.

GUI Equivalent

Configure Analysiswindow > HOW pane

Syntax

-knob | -k <knob-name>=<knob-value>

Arguments

knob-name

An analysis type or collector type may have one or more configuration options (knobs) that provide additional instructions for performing the specified type of analysis. To use a knob, you must specify the knob name and knob value.

Multiple knob options are allowed and can be followed by additional action-options, as well as global-options, if needed.

knob-value

There are values available for each knob. In most cases this is a Boolean value, so for Boolean knobs, specify <knob-name>=true to enable the knob.

NOTE:

Knob behavior may vary depending on the analysis type or collector type.

<knob-name>

Description

accurate-cpu-time-detection=true | false (Windows only)

Default: true

Collect more accurate CPU time data. This option requires additional disk space and post-processing time. Administrator privileges are required.

Supported analysis: runss

analyze-loops=true | false

Default: false

Extend loop analysis to collect advanced loops information such as instruction set usage and display analysis results by loops and functions.

Supported analysis: runss, runsa

analyze-mem-objects=true | false

Default: false

Enable the instrumentation of memory allocation/de-allocation and map hardware events to memory objects. This option is supported only for Linux targets which run on the Intel microarchitectures code named Haswell (or newer).

Supported analysis: memory-access

analyze-openmp=true | false

Default: true for the HPC Performance Characterization analysis; false for other analysis types.

Instrument the OpenMP* runtimes in your application to group performance data by regions/work-sharing constructs and detect inefficiencies such as imbalance, lock contention, or overhead on performing scheduling, reduction, and atomic operations. Using this option may cause higher overhead and increase the result size.

Supported analysis: hotspots, threading, hpc-performance, memory-access, uarch-exploration, runsa

analyze-persistent-memory=true | false

Default: false

Collect performance information for Intel® Optane™ Persistent Memory modules.

Supported analysis: platform-profiler

analyze-power-usage=true | false

Default: false

Collect information about energy consumed by CPU, DRAM, and discrete GPU.

Supported analysis: gpu-hotspots,gpu-offload

analyze-throttling-reasons=true | false

Default: false

Collect information about factors that cause the CPU to throttle.

Supported analysis: system-overview

analyze-xelink-usage=true | false

Default: false

Collect information about data traffic between GPU interconnects (Xe Link) in multi-GPU analysis.

Supported analysis: gpu-hotspots,gpu-offload

atrace-config=<event>

Available events are gfx, input, view, webview, wm, am, audio, video, camera, hal, res, dalvik.

Collect Android framework events from Systrace*.

Supported analysis: runsa

characterization-mode=overview | global-local-accesses | compute-extended | full-compute | instruction-count

Default: overview

Monitor the Render and GPGPU engine usage (Intel Graphics only), identify which parts of the engine are loaded, and correlate GPU and CPU data.

The Characterization mode uses platform-specific presets of the GPU metrics. All presets, except for the instruction-count, collect data about execution units (EUs) activity: EU Array Active, EU Array Stalled, EU Array Idle, Computing Threads Started, and Core Frequency; and each one introduces additional metrics:

  • overview metric set includes additional metrics that track general GPU memory accesses such as Memory Read/Write Bandwidth, GPU L3 Misses, Sampler Busy, Sampler Is Bottleneck, and GPU Memory Texture Read Bandwidth. These metrics can be useful for both graphics and compute-intensive applications.
  • global-local-accesses metric group includes additional metrics that distinguish accessing different types of data on a GPU: Untyped Memory Read/Write Bandwidth, Typed Memory Read/Write Transactions, SLM Read/Write Bandwidth, Render/GPGPU Command Streamer Loaded, and GPU EU Array Usage. These metrics are useful for compute-intensive workloads on the GPU.
  • compute-extended metric group includes additional metrics targeted only for GPU analysis on the Intel processor code name Broadwell and higher. For other systems, this preset is not available.
  • full-compute metric group is a combination of the overview and global-local-accesses event sets.
  • instruction-count metric group counts the execution frequency of specific classes of instructions.

Supported analysis: gpu-hotspots, graphics-rendering, runsa

chipset-event-config="event1,event2 ,..."

Specify a comma-separated list of Android chipset events (up to 5 events) to monitor with the hardware event-based sampling collector.

Supported analysis: runsa

source-analysis=bb-latency | mem-latency

Default value: bb-latency

Collect data on performance-critical basic blocks and issues caused by memory accesses in the GPU kernels. Choose one of the following modes:

  • bb-latency mode helps you identify issues caused by algorithm inefficiencies. In this mode, Intel® VTune™ Profiler measures the execution time of all basic blocks. Basic block is a straight-line code sequence that has a single entry point at the beginning of the sequence and a single exit point at the end of this sequence. During post-processing, VTune Profiler calculates the execution time for each instruction in the basic block. So, this mode helps understand which operations are more expensive.
  • mem-latency mode helps identify latency issues caused by memory accesses. In this mode, Intel® VTune™ Profiler profiles memory read/synchronization instructions to estimate their impact on the kernel execution time. Consider using this option, if you ran the gpu-hotspots analysis in the Characterization mode, identified that the GPU kernel is throughput or memory-bound, and want to explore which memory read/synchronization instructions from the same basic block take more time.

Supported analysis: gpu-hotspots

collect-bad-speculation=true | false

Default value: true

Collect the minimum set of data required to compute top-level metrics and all Bad Speculation sub-metrics.

Supported analysis: uarch-exploration, runsa

collect-core-bound=true | false

Default: false

Collect the minimum set of data required to compute top-level metrics and all Core Bound sub-metrics.

Supported analysis: uarch-exploration, runsa

collect-frontend-bound=true | false

Default value: true

Collect the minimum set of data required to compute top-level metrics and all Front-End Bound sub-metrics.

Supported analysis: uarch-exploration, runsa

collect-cpu-gpu-bandwidth=true | false

Default: false

Collect DRAM bandwidth data for all hosts. Additionally, collect PCIe bandwidth for supported server hosts (Intel® micro-architectures code named Ice Lake and Sapphire Rapids). To view collected data in GUI, enable the Analyze CPU host-GPU bandwidth option.

Supported analysis:gpu-offload

collect-cpu-gpu-pci-bandwidth=true | false

Default: false

Collect PCIe bandwidth for supported server hosts (Intel® micro-architectures code named Ice Lake and Sapphire Rapids). This knob is available for custom analyses only. To view collected data in GUI, enable the Analyze CPU host-GPU bandwidth option.

Supported analysis:runsa

collect-io-waits=true | false

Default: false

Analyze the percentage of time each thread and CPU spends in I/O wait state.

Supported analysis: runsa

collect-memory-bandwidth=true | false

Default: depends on analysis type

Collect data to identify where your application is generating significant bandwidth to DRAM. To view collected data in GUI, enable the Analyze memory bandwidth option.

Supported analysis: performance-snapshot, uarch-exploration, hpc-performance, gpu-hotspots,runsa

collect-memory-bound=true | false

Default value: true

Collect the minimum set of data required to compute top-level metrics and all Memory Bound sub-metrics.

Supported analysis: uarch-exploration, hpc-performance

collect-programming-api=true | false

Default for gpu-hotspots: true, for runss: false.

Analyze execution of SYCL apps, OpenCL™ kernels and Intel® Media SDK programs on Intel HD Graphics and Intel® Iris® Graphics. This option may affect the performance of your application on the CPU side.

Supported analysis: gpu-hotspots, gpu-offload, runsa

collect-retiring=true | false

Default value: true

Collect the minimum set of data required to compute top-level metrics and all Retiring sub-metrics.

Supported analysis: uarch-exploration, runsa

collecting-mode=hw-tracing | hw-tracing

Default value: hw-sampling

Specify the system-wide collection mode to either explore CPU, GPU, and I/O resources utilization with the default event-based sampling mode, or enable the low-overhead hardware tracing and identify a root cause of latency issues.

Supported analysis: system-overview, runsa

computing-tasks-of-interest=computing_task_name[#start_idx#step#stop_idx]

Specify a comma-separated list of GPU computing task names and invocations. Use a search string, if necessary (* and . are supported).

On Windows OS, Intel® VTune™ Profiler does not demangle C++ kernel names during runtime. Instead of searching for the exact C++ kernel name(s), use the search string.

For example, when you set -knob computing-tasks-of-interest=gemm#1#1#4294967295, the search covers all kernels in the source code which have gemm in their name.

Invocations happen in this format:

computing_task_name[#start_idx#step#stop_idx] Default value:*#1#1#4294967295
  • computing_task_name is the name of the GPU computing task (default value is *);
  • start_idx is the number of the first invocation to be profiled (default value is 1);
  • step_idx is the number of the step idx invocation (default value is 1);
  • stop_idx is the number of the last invocation to be profiled (default value is 4294967295, UNIT_MAX)

Supported analysis: gpu-hotspots, runsa

counting-mode=true | false

Default: false

Choose between collecting detailed context data for each PMU event (such as code or hardware context) or the counts of events. Counting mode introduces less overhead but gives less information.

Supported analysis: runsa

cpu-samples-mode=off | stack | nostack

Default: false

Enable to periodically sample the application. Samples can be collected with or without stacks.

Supported analysis: runss

dpdk=true | false

Default: false

Profile DPDK IO API.

Supported analysis: io

dram-bandwidth-limits=true | false

Default: true for the HPC Performance Characterization and Microarchitecture Exploration analysis with collect-memory-bandwidth knob enabled; true for the Memory Access and Microarchitecture Exploration analysis.

Evaluate maximum achievable local DRAM bandwidth before the collection starts. This data is used to scale bandwidth metrics on the timeline and calculate thresholds.

Supported analysis: performance-snapshot, memory-access, uarch- exploration, hpc-performance, runsa

enable-characterization-insights=true | false

Get additional performance insights such as the efficiency of hardware usage, and learn next steps.

Supported analysis: gpu-offload

enable-context-switches=true | false

Default: false

Analyze detailed scheduling layout for all threads in your application, explore time spent on a context switch and identify the nature of context switches for a thread (preemption or synchronization).

Supported analysis: runsa

enable-driverless-collection=true | false

Default: false

Enable driverless Linux Perf collection when possible.

Supported analysis: runsa

enable-gpu-usage=true | false

Default: false

Analyze frame rate and usage of Intel HD Graphics and Intel® Iris® Graphics engines and identify whether your application is GPU or CPU bound.

Supported analysis: runss, runsa

enable-interrupt-collection=true | false

Default: false

Collect interrupt events that alter a normal execution flow of a program. Such events can be generated by hardware devices or by CPUs. Use this data to identify slow interrupts that affect your code performance.

Supported analysis: system-overview.

enable-parallel-fs-collection=true | false

Default: false

Analyze Lustre* file system performance statistics, including Bandwidth, Package Rate, Average Packet Size, and others.

Supported analysis: runsa

enable-stack-collection=true | false

Default: false

Enable Hardware Event-based Sampling Collection with Stacks.

Supported analysis: hotspots, hpc-performance, gpu-offload, runsa

enable-system-cswitch=true | false

Default: false

Analyze detailed scheduling layout for all threads on the system and identify the nature of context switches for a thread (preemption or synchronization).

Supported analysis: runsa

enable-thread-affinity=true | false

Default: false

Analyze thread pinning to sockets, physical cores, and logical cores. Identify incorrect affinity that utilizes logical cores instead of physical cores and contributes to poor physical CPU utilization.

NOTE:

Affinity information is collected at the end of the thread lifetime, so the resulting data may not show the whole issue for dynamic affinity that is changed during the thread lifetime.

enable-user-sync=true | false

Default: false

Collect synchronization data via the User-Defined Synchronization API.

Supported analysis: threading, runss

enable-user-tasks=true | false

Default: false

Analyze tasks, events and counters specified in your application via the Task API. This option causes higher overhead and increases result size.

Supported analysis: hotspots, threading, uarch-exploration, runss, runsa

event-config=<event_name1>,<event_name2>,...

Configure PMU events to collect with the hardware event-based sampling collector. Multiple events can be specified as a comma-separated list (no spaces).

NOTE:

To display a list of events available on the target PMU, enter:

vtune -collect-with runsa -knob event-config=? <target>

The command returns names and short descriptions of available events. For more information on the events, use Intel Processor Events Reference.

Supported analysis: runsa

event-mode=all | user | os

Default: all

Limit event-based sampling collection to OS or USER mode.

Supported analysis: runsa

ftrace-config=<event_name>

Available events are freq, idle, sched, disk, filesystem, irq, kvm, workq, softirq, sync.

Default for Linux targets: sched,freq,idle,workq,irq,softirq

Default for Android targets: sched,freq,idle,workq,filesystem, irq,softirq,sync,disk

Collect Linux Ftrace* framework events.

Supported analysis: runsa, runss

gpu-sampling-interval=<number> between 0.1 and 1000ms

Default: 1.

Specify an interval between GPU samples (in milliseconds).

Supported analysis: gpu-hotspots, graphics-rendering, runss, runsa

io-mode=off | stack | nostack

Default: off

Enable to identify where threads are waiting or compute thread concurrency. The collector instruments APIs, which causes higher overhead and increases result size.

Supported analysis: runss, runsa

ipt-regions-to-load=<number> between 10 and 5000

Default: 1000

Specify the maximum number (10-5000) of code regions to load for detailed analysis.

Supported analysis: anomaly-detection

kernel-stack=true | false

Default: true

Profile system disk IO API.

Supported analysis: io

max-region-duration=<number> between 0.001 and 1000 ms

Default: 100

Specify the maximum duration (0.001-1000ms) of analysis per code region.

Supported analysis: anomaly-detection

mem-object-size-min-thres=<number>

Default: 1024 bytes

Specify a minimal size of memory allocations to analyze. This option helps reduce runtime overhead of the instrumentation.

This option is supported only for Linux targets which run on Intel microarchitectures code named Haswell (or later).

Supported analysis: memory-access

metrics_set=NOC

Default: NOC

Specify the type of metrics set to collect.

Supported analysis: npu

mrte-type=java,dotnet | java,dotnet,python | python

Default: java,dotnet

Specify a type of managed runtime to analyze. Available values: combined .NET* and Java* analysis, combined Java, .NET and Python* analysis, and Python only.

Supported analysis: runss, runsa

no-altstack=true | false

Default: false

Disable using alternative stacks for signal handlers. Consider this option for profiling standard Python 3 code on Linux.

Supported analysis: runss

pmu-collection-mode=detailed | summary

Default: detailed

Choose the detailed sampling-based collection mode to view data breakdown per function and other hotspots. Use the summary counting-based mode for an overview of the whole profiling run. This mode has a lower collection overhead and fast post-processing time.

Supported analysis: uarch-exploration

profiling-mode=characterization (default), code-level-analysis, query-based, time-based

Select a profiling mode for these analyses:

  • GPU Compute/Media Hotspots analysis: Characterize GPU performance issues based on GPU hardware metric presets

  • Custom analysis: Enable a source analysis to identify basic blocks latency due to algorithm inefficiencies or memory latency due to memory access issues

  • NPU Exploration analysis: Select a data collection mode (time-based or query-based)

Supported analysis: gpu-hotspots, runsa, npu

sampling-interval=<number>

For user-mode sampling and tracing types: a number (in milliseconds) between 1 and 1000. Default: 10

For hardware event-based sampling types: a number (in milliseconds) between 0.01 and 1000. Default: 1.

For NPU exploration: a number (in milliseconds) between 0.1 and 1000.

Specify a sampling interval (in milliseconds) between CPU samples. For NPU Exploration analysis, specify the sampling interval for data collection in time-based mode.

Supported analysis: hotspots,runss, threading, ,runsa, system-overview, memory-access, hpc-performance, npu

sampling-mode=sw | hw

Default: sw

Specify a profiling mode.

Use sw to identify CPU hotspots and explore a call flow of your program. This mode does not require sampling drivers to be installed but incurs more collection overhead.

Use hw to identify application hotspots based on such basic hardware events as Clockticks and Instructions Retired. This is a low-overhead collection mode but it requires the sampling driver to be installed on your system.

Supported analysis: hotspots, threading

signals-mode=off | objects | stack | nostack

Default: off

Enable to view synchronization transitions in the timeline and signalling call stacks for associated waits. The collector instruments signalling APIs, which causes higher overhead and increases result size.

Supported analysis: runss

spdk=true | false

Default: false

Profile SPDK IO API.

Supported analysis: io

stack-size=<number>

A number between 0 and 2147483647. Default is 0 (unlimited stack size).

Reduce the collection overhead and limit the stack size (in bytes) processed by the VTune Profiler.

Supported analysis: runsa

stack-stitching=true | false

Default: true

For Intel® oneAPI Threading Building Blocks(oneTBB )-based applications, restructure the call flow to attach stacks to a point introducing a parallel workload.

Supported analysis: runss

stack-type=software | lbr

Default: software

Choose between software stack and hardware LBR-based stack types. Software stacks have no depth limitations and provide more data while hardware stacks introduce less overhead. Typically, software stack type is recommended unless the collection overhead becomes significant. Note that hardware LBR stack type may not be available on all platforms.

Supported analysis: runsa

stackwalk-mode=online | offline

Default: offline

Choose between online (during collection) and offline (after collection) modes to analyze stacks. Offline mode reduces analysis overhead and is typically recommended.

Supported analysis: runss

target-gpu= <domain:bus:device.function[:stack]>

Default: All GPU devices

Select at least one target GPU adapter or stack to collect GPU profiling data. If unset, VTune Profiler collects profiling data for all stacks of all GPUs. If you select a device and do not specify a stack, VTune Profiler collects data for all stacks of the device.

Example: target-gpu=0:58:0.0:1,0:154:0.0

Supported analysis: gpu-offload, gpu-hotspots

uncore-sampling-interval=<number>

For hardware event-based sampling types: a number (in milliseconds) between 1 and 1000. Default: 10.

Specify an interval (in milliseconds) between uncore event samples.

Supported analysis: runsa

waits-mode=off | stack | nostack

Default: off

Enable to identify where threads are waiting or compute thread concurrency. The collector instruments APIs, which causes higher overhead and increases result size.

Supported analysis: runss

Actions Modified

collect, collect-with

Description

Use the knob action-option to configure knob settings for a collect (predefined analysis types) or collect-with (custom analysis types) action where the analysis type supports one or more knobs. Each analysis type or collector type supports a specific set of knobs, and each knob requires a value. In most cases the knob value is Boolean, so you would use True to enable the knob.

To see all knobs available for a predefined analysis type:

vtune -help collect <analysis_type>

To see knobs for a custom analysis type:

vtune -help collect-with <analysis_type>

Example

This example returns a list of knobs for the Threading analysis type:

vtune -help collect threading

This example runs a custom event-based sampling data collection on an Android system enabling collection of Android framework and chipset events.

vtune -collect-with runss -target-system=android -knob sampling-interval=2 -knob cpu-samples-mode=stack -knob ftrace-config=gfx,dalvik -knob chipset-event-config="GMCH_PARTIAL_WR_DRAM.ANY,GMCH_CORE_CLKS" --target-process com.intel.tbb.example.tachyon

This example configures and runs a custom event-based sampling data collection with the stack size limited to 8192 bytes:

vtune -collect-with runsa -knob enable-stack-collection=true -knob stack-size=8192 -knob enable-call-counts=true -knob event-config=CPU_CLK_UNHALTED.REF_TSC:sa=1800000,CPU_CLK_UNHALTED