在服务器上使用vtune

阅读量: searchstar 2021-09-19 15:57:05

Categories： Tags：

¶ 安装

在这里选择合适的下载方式：https://software.intel.com/content/www/cn/zh/develop/tools/oneapi/base-toolkit/download.html

如果没有GUI，就用Command Line Installation里的命令来下载安装。我的是这样的：

wget https://registrationcenter-download.intel.com/akdlm/irc_nas/17769/l_BaseKit_p_2021.2.0.2883.sh
sudo bash l_BaseKit_p_2021.2.0.2883.sh

注意在命令行下安装时，终端窗口要足够大，不然显示不全。

¶ 跑样例工程

安装完成后可以使用vscode来连接服务器使用vtune。官方教程：https://software.intel.com/content/www/us/en/develop/documentation/get-started-with-intel-oneapi-base-linux/top/run-a-sample-project-with-vscode.html

下面简单介绍一下基础用法。

首先安装两个插件：Sample Browser for Intel oneAPI Toolkits和Environment Configurator for Intel oneAPI Toolkits

然后按Ctrl+Shift+P，输入Intel oneAPI，点击Intel oneAPI: Initialize environment variables，一般可以自动发现安装目录（如果安装在默认的/opt/intel/oneapi下的话）。

然后点击最左边导航栏里的oneapi图标，会提示oneapi-cli找不到，问你是否下载，点击Yes即可。但是可能会下载失败。这时可能要在https://github.com/intel/oneapi-cli手动下载下来上传到服务器，然后更名为oneapi-cli放到/usr/local/bin里，记得要给执行权限。

然后选择cpp->Get Started->Base: Vector Add，点击右边的加号，编辑路径后，vscode会新建一个样例工程并且在新窗口打开。可以查看README.md，里面有这个样例工程的介绍以及运行流程。对于Vector Add这个样例，流程是：

先Ctrl+Shift+P跑Intel oneAPI: Initialize environment variables，然后新建终端，执行：

make all
./vector-add-buffers

Running on device: Intel(R) Xeon(R) Platinum 8269CY CPU @ 2.50GHz
Vector size: 10000
[0]: 0 + 0 = 0
[1]: 1 + 1 = 2
[2]: 2 + 2 = 4
...
[9999]: 9999 + 9999 = 19998
Vector add successfully completed on device.

¶ vtune

vtune的教程：https://software.intel.com/content/www/us/en/develop/documentation/get-started-with-vtune/top.html

服务器上用的话，主要看命令行界面的用法：https://software.intel.com/content/www/us/en/develop/documentation/vtune-help/top/command-line-interface/command-syntax.html

注意如果是开普通终端，而不是在vscode里用的话，要先

source /opt/intel/oneapi/setvars.sh

可以运行vtune --help查看基本用法，运行vtune -help collect查看可以搜集的性能数据。

例子：

vtune -collect hotspots -result-dir hotspots -quiet ./vector-add-buffers

注意这个./不能省略。-collect可以缩写为-c，-result-dir可以缩写为-r，-quiet可以缩写为-q。

输出：

vtune: Warning: Hardware collection of CPU events is not possible on this system. Microarchitecture performance insights will not be available.
Running on device: Intel(R) Xeon(R) Platinum 8269CY CPU @ 2.50GHz
Vector size: 10000
[0]: 0 + 0 = 0
[1]: 1 + 1 = 2
[2]: 2 + 2 = 4
...
[9999]: 9999 + 9999 = 19998
Vector add successfully completed on device.

vtune: Warning: Cannot locate debugging information for file `/lib/x86_64-linux-gnu/libgcc_s.so.1'.

vtune: Warning: Cannot locate debugging information for file `/opt/intel/oneapi/vtune/2021.2.0/lib64/runtime/libittnotify_collector.so'.

vtune: Warning: Cannot locate debugging information for file `/opt/intel/oneapi/compiler/2021.2.0/linux/lib/libOpenCL.so.1'.

vtune: Warning: Cannot locate debugging information for file `/opt/intel/oneapi/vtune/2021.2.0/lib64/libtpsstool.so'.

vtune: Warning: Cannot locate debugging information for file `/usr/lib/x86_64-linux-gnu/libstdc++.so.6'.

vtune: Warning: Cannot locate debugging information for file `/lib64/ld-linux-x86-64.so.2'.
Elapsed Time: 1.510s
    CPU Time: 1.450s
        Effective Time: 1.450s
            Idle: 0.030s
            Poor: 1.420s
            Ok: 0s
            Ideal: 0s
            Over: 0s
        Spin Time: 0s
            Imbalance or Serial Spinning: 0s
            Lock Contention: 0s
            Other: 0s
        Overhead Time: 0s
            Creation: 0s
            Scheduling: 0s
            Reduction: 0s
            Atomics: 0s
            Other: 0s
    Total Thread Count: 4
    Paused Time: 0s

Top Hotspots
Function                                   Module                        CPU Time
-----------------------------------------  ----------------------------  --------
clCreateContext                            libintelocl.so                  0.475s
Intel::OpenCL::Utils::OclDynamicLib::Load  libcpu_device.so.2021.11.3.0    0.178s
memcmp                                     libc-dynamic.so                 0.155s
OS_BARESYSCALL_DoCallAsmIntel64Linux       libc-dynamic.so                 0.087s
_GLOBAL__sub_I_acl_platform.cpp            libalteracl.so                  0.060s
[Others]                                   N/A                             0.494s

Top Tasks
Task Type         Task Time  Task Count  Average Task Time
----------------  ---------  ----------  -----------------
tbb_parallel_for     0.086s          11             0.008s
tbb_custom           0.033s           8             0.004s
Collection and Platform Info
    Application Command Line: ./vector-add-buffers 
    Operating System: 4.15.0-144-generic DISTRIB_ID=Ubuntu DISTRIB_RELEASE=18.04 DISTRIB_CODENAME=bionic DISTRIB_DESCRIPTION="Ubuntu 18.04.5 LTS"
    Computer Name: pubchain-dev
    Result Size: 7.2 MB 
    Collection start time: 06:18:08 19/09/2021 UTC
    Collection stop time: 06:18:10 19/09/2021 UTC
    Collector Type: User-mode sampling and tracing
    CPU
        Name: Intel(R) Xeon(R) Processor code named Cascadelake
        Frequency: 2.500 GHz 
        Logical CPU Count: 4
        Cache Allocation Technology
            Level 2 capability: not detected
            Level 3 capability: not detected

If you want to skip descriptions of detected performance issues in the report,
enter: vtune -report summary -report-knob show-issues=false -r <my_result_dir>.
Alternatively, you may view the report in the csv format: vtune -report
<report_name> -format=csv.

¶ 查看结果

运行vtune -help report来查看如何生成报告。最简单的例子：

vtune -report hotspots -r <结果目录>

或者生成summary（其实就是刚跑完输出的那个东西）:

vtune -report summary -r hotspots6/

或者把结果目录传送到本地，然后在本地也安装vtune，然后

vtune-gui <结果目录>

或者也可以参考这篇文章：vnc通过ssh隧道连接到Linux服务器，然后直接用服务器上的vtune-gui打开结果文件，并把窗口显示在本地：

在这里插入图片描述
事实上，通过vnc的方式，可以直接建一个vtune工程来做性能测试：

在这里插入图片描述

linux perf得到完整调用关系

perf.data生成火焰图