Profiling Hardware Without Intel Sampling Drivers
Use this collection of recipes to set up Linux* Perf*-based performance profiling with Intel® VTune™ Profiler in the driverless mode. Understand benefits and identify workarounds for possible limitations.
Content expert: Jeffrey Reinemann , Alexey Bayduraev
Intel processors provide performance monitoring unit (PMU) events that you can employ to analyze how effectively your code utilizes hardware resources. VTune Profiler can collect and analyze PMU events for microarchitecture analysis types like:
- HPC Performance Characterization
- Memory Access
- Microarchitecture Exploration
For PMU event-based analysis, VTune Profiler uses Intel sampling drivers that require administrative privileges to install them on a target system. If you do not have administrative privileges or your environment does not allow third-party drivers to be injected to the systems, VTune Profiler cannot access PMU events via Intel sampling drivers and determine hardware performance bottlenecks. For those cases, VTune Profiler has adopted hardware performance monitoring capabilities through a built-in Linux Perf performance monitoring system.
VTune Profiler enables the hardware event-based sampling analysis in the Perf driverless mode when these conditions are met:
Intel sampling drivers cannot be installed (for example, if installed without root privileges).
Collection with stacks is selected with a non-zero (unlimited) stack size and the requirements for driverless collection are satisfied.
The option to use driverless collection is enabled and the requirements for driverless collection are satisfied.
VTune Profiler extends driverless support for Linux Perf. This provides the profiling functionality, a collection overhead, and a trace size competitive with the solution based on using Intel sampling driver. However, VTune Profiler capabilities in the driverless mode depend on your Linux OS configuration. There may be some limitations. The examples in this recipe describe those limitations.
To enable the Perf driverless collection to match all the hardware profiling functionality provided with Intel drivers, you must have administrative privileges to configure the system options described below.
To check the collector type used for the analysis (Perf or Intel sampling driver (SEP)), see the Collection and Platform Info section of the Summary window.
INGREDIENTS:
VTune Profiler - Run in driverless mode after meeting these prerequirements.
Core and uncore events. All hardware event-based collections in VTune Profiler use core PMU events. Some of them such as Memory Access and IO analysis types require access to uncore events that enable collecting metrics like DRAM bandwidth, QPI/UPI bandwidth, PCI bandwidth, and others.
Perf for Linux kernel 2.6.32 and higher. PMU events are exposed by Linux kernel through /sys/bus/event_source/devices/cpu and /sys/bus/event_source/devices/uncore_* directories. Empty directory content may indicate that the system configuration does not support PMU event collection. In this case, either update the OS or install the Intel sampling driver.
NOTE:In hybrid architectures with performance cores (P-Cores) and efficient cores (E-Cores), P-Core events are exposed through /sys/bus/event_source/devices/cpu_core. E-Core events are exposed through /sys/bus/event_source/devices/cpu_atom./proc/sys/kernel/perf_event_paranoid value is equal to or less than 1.
NOTE:When you run the Microarchitecture Exploration analysis on Intel® microarchitecture codenamed Skylake (or Skylake server), you must use Linux kernel version 4.3 or newer because front end event collection is necessary for this purpose. To verify if your system can support this analysis, see if the frontend file located in /sys/devices/cpu/format/frontend.RECIPES FOR LIMITATIONS:
<vtune_install_dir>/bin64/vtune-self-checker.sh
Enable System Wide or User Process Profiling
Analysis types: All analyses
Concepts: System-wide analysis collects performance information about all processes running on the system, including system services and so on.
Driverless mode limitations: Additional configuration is required to enable system-wide or user process profiling.
To enable system-wide analysis in the driverless mode:
Configure a VTune Profiler project and from the WHAT pane select either the Profile System target or the Launch Application target with the Analyze system-wide option enabled.
Check the /proc/sys/kernel/perf_event_paranoid file value with the following command:
cat /proc/sys/kernel/perf_event_paranoid
If the value is less than 1, VTune Profiler can proceed with the system-wide collection.
If the perf_event_paranoid value is equal to 1 (which limits the collection to user processes only) or more than 1 (which prevents VTune Profiler from using the Perf driverless mode), set the perf_event_paranoid value to 0 for the system-wide collection:
echo 0 > /proc/sys/kernel/perf_event_paranoid
In some environments, perf_event_paranoid is regulated by the security policy. For more information about Linux Perf security requirements, see https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html.
Intel sampling driver limitations: By default, the Intel sampling driver supports system-wide collection. But if it is built and loaded with the --per-user option, the collection will be limited to user processes only.
Enable Core and Uncore Event Collection
Analysis types: Memory Access, HPC Performance Characterization, and other analysis types based on uncore events
Core events can be collected both system wide and per user processes. To collect uncore events in the Perf driverless mode, enable system-wide analysis.
Driverless mode limitations:
Memory Access analysis requires access to uncore events and will not run without ability to collect them. Other analysis types, like HPC Performance Characterization, will run but miss metrics based on uncore events such as DRAM Bandwidth, OPA Interconnect Bandwidth, and Packet Rate.
Uncore collection in the driverless mode is not supported on Intel Atom® processors.
To collect uncore events in the driverless mode:
Set the perf_event_paranoid value to 0 to enable system-wide performance monitoring, which is a prerequisite for the uncore event collection.
Intel sampling driver limitations: None
Enable Multi-Process Profiling
Analysis types: all
By default, the Linux kernel limits the size of the memory available for capturing performance data by 518Kb. To know the current value, enter:
cat /proc/sys/kernel/perf_event_mlock_kb
Driverless mode limitations: For some parallel applications (for example, MPI applications with multiple ranks per node) in the user process collection mode on multi-core systems (>64 logical cores), the limit of 518Kb tends to be reached and the data collection will not be available.
To enable multi-process profiling on a multi-core system:
Set the perf_event_paranoid value to 0 to enable system-wide performance monitoring.
Intel sampling driver limitations: none. Any number of processes with default settings can be profiled.
Profile a Large Number of PMU Events on Multi-Core Systems
Analysis types: Microarchitecture Exploration
Driverless mode limitations: Linux Perf allocates file descriptors for every configured PMU event on each CPU. So, on a multi-core system with a long events list used by such analyses as Microarchitecture Exploration, this limit is easily reachable and can prevent the collection in the driverless mode.
To support profiling a large number of PMU events in the driverless mode:
Check the limit of opened files:
ulimit -n
If required, increase the limit in the /etc/security/limits.conf file. To do this, you must have administrator privilege. Increase the limit by adding or changing these lines (particular numbers are chosen as examples):
* soft nofile 65535 * hard nofile 65535
- If you increased the limit in step 2, log out of the shell or close it and reopen a secure shell connection. Log back in.
NOTE:With administrator privilege, you can set the limit for a specific user. The change should be visible when the user logs in again.
For more information on using the limits.conf file, see http://man7.org/linux/man-pages/man5/limits.conf.5.html.
Intel sampling driver limitations: None.
Enable Stack Sampling
Analysis types: Hotspots (Hardware Event-Based Sampling mode), Threading (Hardware Event-Based Sampling and Stack Stitching mode), HPC Performance Characterization (Collect stacks option enabled), GPU Compute/Media Hotspots (Collect stacks option enabled).
Driverless mode limitations:
Default 1024 byte stack size may not be enough for a full stack unwinding if a function intensively allocates data on the stack. This may lead to [Skipped stack frame(s)] displayed in the collected data.
Linux kernel versions older than 3.7 support only frame-pointer (FP) based stack unwinding. This means that VTune Profiler can provide no stacks for binaries built without frame-pointer (-fomit-frame-pointer compiler option), as well as no Glibc stacks since Glibc is built without frame-pointers.
To avoid issues with stack unwinding in the driverless mode:
Increase the stack size. For example:
vtune -collect hotspots -knob sampling-mode=hw -knob enable-stack-collection=true -knob stack-size=2048 <application>
Otherwise, switch to the Intel sampling driver and set the Stack size option to 0 (unlimited value).
Stack sampling collection with the Intel sampling driver depends on the kernel implementation. The collection requires an update for a new kernel version, which may bring additional product maintenance cost. To reduce this cost, VTune Profiler uses the driverless mode for all analysis types with stacks collection enabled, even when the Intel sampling driver is loaded. If you need to switch to the Intel sampling driver for stack sampling collection, do one of the following:
- Create a custom analysis type and disable the Enable driverless collection option
- Use the corresponding command line configuration
vtune -collect-with runsa -knob enable-driverless-collection=false -knob event-config=<event-list> <application>
Intel sampling driver limitations: No limitation for the stack unwinding since the Intel sampling driver uses a different algorithm of call stack collection. The driver may require an update if your kernel version is newer than the latest kernel version supported by VTune Profiler.
Collect Context Switches
Analysis types: Threading
Concepts: Context switch collection helps expose metrics based on thread Inactive Wait time resulted from either synchronization or thread preemption.
Driverless mode limitations: Linux Perf collects context switches from kernel version 4.3 and higher. Identification of the context switch reason (synchronization or preemption) is available from kernel version 4.17. For older kernel versions, VTune Profiler switches the collection to the Intel sampling driver if it is available on the system.
Intel sampling driver limitations: none.
Resolve Symbols for Kernel Functions
Analysis types: All
Driverless mode limitations: Additional manual configuration of the kptr_restrict file is required.
To associate performance data with kernel function names:
Set the kptr_restrict configuration file value to 0 as a system administrator:
echo 0 > /proc/sys/kernel/kptr_restrict
Setting the value to 1 limits the file name resolution to user-level modules.
Intel sampling driver limitations: None. The Intel driver resolves kernel symbols if the /boot/System.map-<kernel_version> file is accessible for reading or /proc/sys/kernel/kptr_restrict is set to 0.
Avoid Resource Contention with the NMI Watchdog
Analysis types: all
Driverless mode limitations: NMI watchdog (a hard lockup detector) utilizes one CPU performance counter register that becomes unavailable for Linux Perf. This can increase the number of multiplexing groups and, as a result, impact the accuracy of statistical sampling data.
To improve the accuracy of analysis runs with long events lists in the driverless mode:
Disable the NMI watchdog, using administrative privileges:
echo 0 > /proc/sys/kernel/nmi_watchdog
When the driverless Perf collection is complete, you can re-enable the NMI watchdog (using the administrative privileges):
echo 1 > /proc/sys/kernel/nmi_watchdog
Intel sampling driver limitations: none. Intel driver automatically stops the NMI watchdog for the collection time to avoid this problem with data accuracy.
Reduce Collection Overhead
Analysis types: all
Driverless mode limitations: Linux Perf collection may incur an overhead on CPU intensive applications, since it fully loads all CPUs.
To reduce the collection overhead in the driverless mode:
To reduce the trace size for stack sampling collections, the VTune Profiler uses a Linux Perf trace compression, which may introduce an additional overhead. To avoid this, disable the trace compression with the -run-pass-thru option:
vtune -collect hotspots -knob sampling-mode=hw -knob enable-stack-collection=true -run-pass-thru=--perf-compression=0 <application>
This can reduce collector overhead in rare cases, but the trace size increases dramatically.
In some real-time and telecom applications, the default per-CPU trace collection mode can cause collection overhead. To overcome this, disable the per-CPU trace collection mode with the -run-pass-thru option:
vtune -collect hotspots -knob sampling-mode=hw -knob enable-stack-collection=true -run-pass-thru=--perf-threads=none <application>
Set the limit of CPU time consumption by Linux Perf collector. For example, for a 10% limit, use the following command (with administrative privileges):
echo 10 > /proc/sys/kernel/perf_cpu_time_max_percent
This can drop the sampling frequency and statistical accuracy to reach the limit.
Intel sampling driver limitations: none.
Enable Using Driverless Mode When Required
VTune Profiler uses the Intel sampling driver if it is loaded in all cases except for the stack sampling collection. To make the VTune Profiler use the driverless Perf mode for sampling without stacks, create a custom analysis type and select the Enable driverless collection option in the GUI, or set the command line knob value to enable-driverless-collection=true as follows:
vtune -collect-with runsa -knob enable-driverless-collection=true -knob event-config=<event-list> <application>
The option is available starting with the VTune Amplifier 2019 Update 4.
Enable Profiling Capabilities for the Group
Driverless mode limitations: Setting the perf_event_paranoid option to a lower value could be inappropriate as this option applies to all users. Instead, you could set Linux capabilities to a specific user group and binary.
To enable capabilities for perf tool:
The required capabilities cannot be assigned for a file system mounted with the nosuid option, or if the file system does not support extended file attributes.
For Linux kernel versions older than 5.8, use CAP_SYS_ADMIN. To set up this configuration, run the vtune-set-perf-caps.sh script with this parameter:
vtune-set-perf-caps.sh -v cap_sys_admin
For Linux kernel versions 5.8 and newer, use CAP_PERFMON. To set up this configuration, run the vtune-set-perf-caps.sh script with this parameter:
vtune-set-perf-caps.sh -v cap_perfmon
Alternatively, you can set up this configuration manually:
- Create a vtune group for privileged amplxe-perf users.
- Assign the vtune group to the Perf tool executable.
- Restrict access to the executable to only those users who are in the vtune group.
# cp amplxe-perf amplxe-perf-priv # groupadd vtune # chgrp vtune amplxe-perf-priv # chmod o-rwx amplxe-perf-priv
- Assign the required capabilities to the Perf tool executable.
# setcap -v "cap_perfmon,cap_sys_ptrace,cap_syslog=ep" amplxe-perf-priv # getcap amplxe-perf-priv amplxe-perf-priv = cap_sys_ptrace,cap_syslog,cap_perfmon+ep
If the installed libcap does not support cap_perfmon, use 38 instead:# setcap "38,cap_sys_ptrace,cap_syslog=ep" amplxe-perf-priv # getcap amplxe-perf-priv amplxe-perf-priv = cap_sys_ptrace,cap_syslog,38+ep
For more information on Linux capabilities, see https://man7.org/linux/man-pages/man7/capabilities.7.html.
For more information on perf_event access control, see https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html