Interpreting Data on Concurrency and CPU Usage

Start with analyzing the application-level data provided in the Summary window for this analysis result. Use the Elapsed time as your primary indicator and a baseline for comparison of results before and after optimization.

Explore the CPU Usage and Thread Concurrency histograms that represent the Elapsed time and utilization level for the specified number of running threads and available CPUs. Ideally, your longest bars should be within the Ok or Ideal utilization range defined by the Intel® VTune™ Amplifier .

To identify functions that do not use available processor time effectively, explore the Bottom-up pane.

To identify functions with poor CPU usage, explore the Hotspots by CPU Usage viewpoint. By default, the functions are sorted by Poor processor utilization type. The most critical functions are provided first. You can view the time distribution per processor utilization type by clicking the button at the CPU Time by Utilization column header to expand the column. You may also right-click the header and select Set Column as Data of Interest to use the values from this column for calculations in other views (for example, stack contribution in the Call Stack pane).

To identify functions that ran serially and did not use available cores effectively (functions with poor concurrency), switch to the Hotspots by Thread Concurrency viewpoint. The functions are sorted by CPU time with poor concurrency level. The usage mode is similar to the Hotspots by CPU Usage viewpoint.

You should focus your optimization efforts on functions with the longest poor CPU time (red bars if the bar format is selected). Next search for the longest over-utilized time (blue bars).

The overall goal of optimization is to achieve Ideal (green ) or OK (orange ) utilization and shorten the Poor and Over CPU utilization/concurrency.

VTune Amplifier also measures the Overhead time, which is the CPU time spent on the overhead of known synchronization or threading libraries, such as system synchronization APIs, Intel Threading Building Blocks (Intel TBB), and OpenMP*. If the Overhead time exceeds the threshold set up by Intel architects for your processor type, the VTune Amplifier highlights these values in pink in the Bottom-up/Top-down Tree panes. Hover over the highlighted cell to get performance tuning advice:

The Timeline pane at the bottom of the Bottom-up/Top-down Tree windows shows the thread behavior in your application and how CPU Usage and Thread Concurrency metrics are changing over time. Analyze the data, select the problem area, and zoom in to selection using the context menu options. For example, in the figure below during 16.7 - 18s time range, the CPU usage level is poor. VTune Amplifier calculates the overall CPU Usage metric as the sum of CPU time per each thread of the Threads area. Maximum CPU Usage value is equal to [number of processor cores] x 100%. In the example below, CPU Time for a 4-core system is 100%, which means that the CPUs were not effectively utilized during this time range.

To understand what your application was doing during this time frame, select the thread of interest in the Timeline pane, for example: OMP Worker Thread#1 thread. VTune Amplifier synchronizes your selection with the Bottom-up pane by highlighting the functions corresponding to the selected thread. Identify functions with high CPU time (hotspots) and double-click a hotspot to identify the code lines that caused the issue.

You can identify issues with the call sequences in your application and improve performance by revising the way functions are called. The following methods to locate potential issues are available:

• Top-down Tree pane: Analyze the Total and Self time data for callers and callees of the hotspot function to understand whether this time can be optimized.

• Call Stack pane: Identify the highest contributing stack for the program unit(s) selected in the Bottom-up or Top-down Tree panes. Use the navigation buttons to see the different stacks that called the selected program unit(s). The contribution bar shows the contribution of the currently visible stack to the overall time spent by the selected program unit(s). You can also use the drop-down list in the Call Stack pane to view data for different types of stacks.

When you identified a critical function, double-click it to open the Source/Assembly window and analyze the source code. From the Timeline pane, you can double-click the transition line to open the call site for this transition. You can open the code editor directly from the VTune Amplifier and edit your code (for example, adding parallelism, rebalancing or reducing contention).

• Use the Locks and Waits analysis to understand possible reasons why your application does not use the available processor effectively.

• Run the comparison analysis to understand the performance gain you obtain after your optimization.

• Run an advanced hardware event-based sampling analysis to identify hardware issues affecting the performance of your application.