Start with the

Summary

window that provides a short overview of the target performance and, for Linux* targets, introduces the

I/O Wait Time

metric that helps you estimate whether your application is I/O-bound:

The I/O Wait Time metric represents a portion of time when threads reside in I/O wait state while there are idle cores on the system, and the number of counted threads is not greater than the number of idling cores. This aggregated I/O Wait Time metric is an integral function of I/O Wait metric that is available in the Timeline pane of the Bottom-up view.

Scroll down to the

Disk Input and Output Histogram

and estimate how quickly storage requests are served by the kernel sub-system. Use the

Operation type

drop-down menu to select the type of I/O operation you are interested in. For example, for the

write

type of I/O operations, 2-4 storage requests executed for more than 0.06 seconds are qualified by the

VTune

Profiler

as slow:

To get more details on this type of I/O request, switch to the

Bottom-up

window.

In the Bottom-up window, you may select an area of interest on the timeline, right-click and select the

Zoom In and Filter In by Selection

context menu option. The grid view and context summary histogram will be updated to show the data for the selected time range.

In this example there were 2-4 slow write requests executed at the 6th second of the target execution.

When you zoom in to the area of interest, you have a closer look at all the metrics and understand what caused high I/O Wait time.

For Linux targets, the

VTune

Profiler

collects the

I/O Wait

type of context switches caused by I/O accesses from the thread (slate blue bars in the Thread area) and also provides a system-wide

I/O Wait

metric in the

CPU Activity

area. Use this metric data to identify imbalance between I/O and compute operations. System-wide I/O Wait shows the time when system cores are idle, but there are threads in a context switch caused by I/O access. Use this metric to estimate the performance dependency on the storage medium. For example, 100% value of the

I/O Wait

metric means that all cores of the system are idle, but there are threads (greater or equal than the number of CPU cores) blocked by I/O requests. To solve this problem, consider changing the logic of an application to run compute threads in parallel with I/O tasks. Another alternative is to use faster storage. 0% value of the

I/O Wait

metric means one of the following:

Explore the

I/O Queue Depth

area that shows the number of I/O requests submitted to the storage device. Spikes correspond to the maximum number of requests. Zero-valued gaps on the

I/O Queue Depth

chart correspond to the moments of time when the storage was not utilized at all. You may enable the

Slow

markers for the

I/O Queue Depth

metric to see where exactly slow I/O packets are scheduled for execution:

To identify points of high bandwidth, analyze the

I/O Data Transfer

area. It shows the number of bytes read from or written to the storage device.

For server platforms based on the Intel microarchitecture code name Sandy Bridge EP and later, the

VTune

Profiler

provides the PCIe Bandwidth metrics on the timeline:

Use this data to identify time ranges where your application could be stalled due to approaching bandwidth limits of the PCIe bus. These metrics do not attribute I/O requests to threads/cores/sockets (see Uncore Event Count window in the Hardware Events viewpoint for that).

Correlate slow I/O requests with instrumented user-space activities. For the storage analysis, the

VTune

Profiler

instruments all user-space I/O functions and enables you to view a full call stack pointing to the exact API invocation.

To view a Task Time call stack for a particular I/O API call, select the required

I/O API

marker on the timeline and explore the stack in the

Call Stack

pane: