![pmu-tools] (http://halobates.de/pmu-tools.png)

pmu tools is a collection of tools for profile collection and performance analysis on Intel CPUs on top of Linux perf. This uses performance counters in the CPU.

• toplev now has a (draft) tutorial and manual
• toplev now detects kernel counter scheduling problems based on the kernel version and automatically disables nodes with unsupported events. Can be overridden with --force-events.
• toplev now defaults to measuring the whole system even on non SMT. The old mode can be still enabled with --single-thread. It should be only used on single threads and when the other thread is idle on SMT. In addition there is a new --core option to select the CPUs to measure.
• toplev can now automatically sample workloads with --run-sample
• Added cputop utility to easily enable/disable hyper threading
• toplev updated to TopDown 2.9:
  • Many fixes to SMT support. SMT now supported on Haswell.
  • Many bug fixes to metrics
  • Initial Haswell Server support
  • Add a Sandy Bridge EP model
  • Lots of new metrics and nodes: e.g. Core IPC, MUX confidence, BPTB (Branch per Taken Branch), SMT_2T utilization, IFetch_Line_utilization
  • Initial Broadwell model support
  • Improve sample event support. Now enable them by default.
  • --per-core and --per-socket support dropped with SMT
• The new tl-serve.py tool displays toplev output in a web browser.

• The "ocperf" wrapper to "perf" that provides a full core performance counter event list for common Intel CPUs. This allows to use all the Intel events, not just the builtin events of perf. Can be also used as a library from other python programs
• The "toplev.py" tool to identify the micro-architectural bottleneck for a workload. This implements the TopDown or TopDown2 methology.
• The "ucevent" tool to manage and compute uncore performance events. Uncore is the part of the CPU that is not core. Supports many metrics for power management, IO, QPI (interconnect), caches, and others. ucevent automatically generates event descriptions for the perf uncore driver and pretty prints the output. It also supports computing higher level metrics derived from multiple events.
• A library to resolve named intel events (like INST_RETIRED.ANY) to perf_event_attr (jevents)
• A library for self profiling with Linux since Linux 3.3. Self profiling is a program monitoring its own execution, either for controlled benchmarking or to optimize itself. For self-profiling on older kernels you can use [simple-pmu] (http://halobates.de/simple-pmu)
• Support for Intel "offcore" events on older Linux systems where the kernel perf subsystem does not support them natively. Offcore events allow to categorize memory accesses that go outside the core.
• Workarounds for some issues with offcore events on Sandy Bridge EP (Intel Xeon E5 v1) This is automatically enabled for the respective events with ocperf, and also available as a standalone program or python library.
• A variety of tools for plotting and post processing perf stat -I1000 -x, or toplev.py -I1000 -x, interval measurements.

• An example program for address profiling on Nehalem and later Intel CPUs (addr)
• Some utility programs to access pci space or msrs on the command line
• A utility program to program the PMU directly from user space (pmumon.py) for counting. This is mainly useful for testing and experimental purposes.
• A program to print the currently running events (event-rmap)
• Support for analyzing the raw PEBS records with perf.
• A pandas/scipy data model for perf.data analytics (work in progress)

Check out the repository. Run the tools from the directory you checked out (but it does not need to be the current directory) They automatically search for other modules and data files in the same directory the script was located in.

You can set the PATH to include the repository to run the tools from other directories:

export PATH=/path/to/pmu-tools:$PATH

You want to:

• understand CPU bottlenecks on the high-level: use toplev.
• display toplev output graphically: use tl-server or toplev --graph
• know what CPU events to run, but want to use symbolic names: use ocperf.
• measure interconnect/caches/memory/power management on Xeon E5+: use ucevent
• Use perf events from a C program: use jevents, self, addr
• Query CPU topology or disable HyperThreading: use cputop
• Change Model Specific Registers: use msr
• Change PCI config space: use PCI

The other tools are for more obscure usages.

All tools (except for parser/) should work with a python 2.7 standard installation. All need a reasonably recent perf (RHEL5 is too old) ocperf.py should work with python 2.6, or likely 2.5 when the json module is installed.

Except for the modules in parser/ there are no special dependencies outside a standard python install on a recent Linux system with perf.

old. parser/ needs a scipy stack with pandas and pyelftools.

The perf tool should not be too old.

toplev has kernel dependencies, please see https://github.com/andikleen/pmu-tools/wiki/toplev-kernel-support

ocperf is a wrapper to "perf" that provides a full core event list for common Intel CPUs. This allows to use all the Intel defined events, not just the builtin events of perf.

A more detailed introduction is in Andi's blog

ocperf.py list List all the events perf and ocperf supports on the current CPU

ocperf.py stat -e eventname ... 

ocperf.py record -c default -e eventname ...

ocperf.py report

When a older kernel is used with offcore events (events that count types of memory accesses outside the CPU core) that does not support offcore events natively, ocperf has to run as root and only one such profiling can be active on a machine.

When "-c default" is specified for record, the default sampling overflow value will be filled in for the sampling period. This option needs to be specified before the events and is not supported for all CPUs. By default perf uses a dynamic sampling period, which can cause varying (and sometimes large) overhead. The fixed period minimizes this problem.

If you have trouble with one of the many acronyms in the event list descriptions, the Intel optimization manual describes many of them.

ocperf.py can be also used as a python module to convert or list events for the current CPU:

	import ocperf

	emap = ocperf.find_emap()
	if not emap:
		sys.exit("Unknown CPU or cannot find event table")
	ev = emap.getevent("BR_MISP_EXEC.ANY")
	if ev:
		print "name:", ev.output()
		print "raw form:", ev.output(use_raw=True)
		print "description:, ev.desc

To retrieve data for other CPUs set the EVENTMAP environment variable to the csv file of the CPU before calling find_emap()

The msr.py, pci.py, latego.py can be used as standalone programs or python modules to change MSRs, PCI config space or enable/disable the workarounds.

For example to set the MSR 0x123 on all CPUs to value 1 use:

$ sudo ./msr.py 0x123 1

To read MSR 0x123 on CPU 0

$ sudo ./msr.py 0x123

To read MSR 0x123 on CPU 3:

$ sudo python
>>> import msr
>>> msr.readmsr(0x123, 3)

To set bit 0 in MSR 0x123 on all CPUs:

$ sudo python
>>> import msr
>>> msr.writemsr(0x123, msr.readmsr(0x123) | 1)

(this assumes the MSR has the same value on all CPUs, otherwise iterate the readmsr over the CPUs)

Identify the micro-architectural bottleneck of a workload.

The bottlenecks are expressed as a tree with different levels (max 5). Each bottleneck is only meaningful if the parent higher level crossed the threshold (it acts similar to a binary search). The tool automatically only prints meaningful ratios, unless -v is specified.

This follows the "Top Down" methology. The best description of the method is in the "A top-down method for performance analysis and counter architecture" paper (ISPASS 2014, available here) I didn't invent it, I'm just implementing it.

A more gentle introduction is in andi's blog

Please also see the manual and tutorial

toplev.py only supports counting, that is it cannot tell you where in the program the problem occurred, just what happened. There is now an experimential --show-sample option to suggest sampling events for specific problems. The new --run-sample option can also automatically sample the program by re-running.

Requires Intel CPUs Sandy Bridge (Core 2nd gen, Xeon 5xxx) or newer or Atom Silvermont or newer. Quark or Xeon Phi are not supported.

By default the simple high level model is used. The detailed model is selected with -lX, with X being the level.

On non-SMT systems only the program is measured by default, while with SMT on the whole system is measured.

[IVB model] (http://halobates.de/ivb-hierarchy.svg) [Simple model] (http://halobates.de/simple-hierarchy.svg)

Usage:

usage: toplev [options] perf-arguments

Estimate on which part of the CPU pipeline a workload bottlenecks using the TopDown model. The bottlenecks are expressed as a tree with different levels.

Examples:

./toplev.py -l2 program measure program in level 2

./toplev.py --all -a sleep X measure whole system for X seconds

./toplev.py -o logfile.csv -x, -p PID measure pid PID, outputting in CSV format

-h, --help            show this help message and exit
--verbose, -v         Print all results even when below threshold
--kernel              Only measure kernel code
--user                Only measure user code
--print-group, -g     Print event group assignments
--no-desc             Don't print event descriptions
--csv CSV, -x CSV     Enable CSV mode with specified delimeter
--interval INTERVAL, -I INTERVAL
                      Enable interval mode with ms interval
--output OUTPUT, -o OUTPUT
                      Set output file
--graph               Automatically graph interval output with tl-barplot.py
--title TITLE         Set title of graph
--xkcd                Use xkcd plotting mode for graph
--level LEVEL, -l LEVEL
                      Measure upto level N (max 5)
--metrics, -m         Print extra metrics
--raw                 Print raw values
--sw                  Measure perf Linux metrics
--tsx                 Measure TSX metrics
--all                 Measure everything available
--frequency           Measure frequency
--no-group            Dont use groups
--no-multiplex        Do not multiplex, but run the workload multiple times
                      as needed. Requires reproducible workloads.
--show-sample         Show command line to rerun workload with sampling
--run-sample          Automatically rerun workload with sampling
--valcsv VALCSV, -V VALCSV
                      Write raw counter values into CSV file
--stats               Show statistics on what events counted
--power               Display power metrics

Other perf arguments allowed (see the perf documentation) After -- perf arguments conflicting with toplevel can be used.

The lower levels of the measurement tree are less reliable than the higher levels. They also rely on counter multi-plexing and cannot use groups, which can cause larger measurement errors with non steady state workloads. Enabling metrics will also force multi-plexing.

(If you don't understand this terminology; it means measurements in higher levels are less accurate and it works best with programs that primarily do the same thing over and over)

If the program is very reproducible it is possible to avoid multiplexing with --no-multiplex. This will run the program many times. The perf -r option (to rerun multiple times to measure the standard deviation) is currently not supported by toplev though.

When using level > 1, it's recommended to measure the work load only after the startup phase by using interval mode or attaching later. level 1 or running without -d is generally the most reliable. The startup phase tends to be multiplexing unfriendly, as it does not execute for long enough.

If your perf stat is new enough (3.12+) the --initial-delay option is useful to skip the startup phase.

Toplev uses many events and is demand of the kernel perf driver. A number of kernel workarounds may be needed. See https://github.com/andikleen/pmu-tools/wiki/toplev-kernel-support

Please see the [ucevent documentation] (http://github.com/andikleen/pmu-tools/tree/master/ucevent/#ucevent-uncore-monitoring)

interval-plot.py can plot the output of perf stat -I1000 -x

Requires matplotlib to be installed.

Below is the level 2 toplev measurement of a Linux kernel compile. Note that tl-barplot below is normally better to plot toplev output.

![plot-example] (http://halobates.de/interval.plot.l2.png)

This converts the output of perf stat -Ixxx -x, / toplev.py -Ixxx -x, to a normalized output (one column for each event). This allows easier plotting and processing with other tools (spreadsheets, R, JMP, gnuplot etc.)

Plot an already normalized data file. Requires pyplot to be installed.

Plot output from toplev.py -I 1000 -v -x, --output file.csv -lLEVELS toplev outputs percentages, so it's better to use a stacked plot, instead of the absolute values interval-plot displays. tl-barplot implements a stacked barplot plot.

Requires matplotlib to be installed.

toplev.py -I 100 -l3 --title "GNU grep" --graph grep -r foo /usr/*

![tl-barplot-example] (http://halobates.de/grep.3.svg)

With a new enough matplotlib you can also enable xkcd mode (install Humor Sans first)

toplev.py -I 100 -l2 --title "povray" --xkcd povray --benchmark

![tl-barplot-xkcd] (http://halobates.de/povray.2.png)

Display toplev.py output in a web browser.

Download dygraphs. Only needs to be done once.

wget http://dygraphs.com/1.0.1/dygraph-combined.js

Run toplev:

toplev.py --all -I 100 -o x.csv ... 
tl-serve.py x.csv

Then browse http://localhost:9001/ in your web browser.

query cpu topology and print all matching cpu numbers cputop "query" ["format"]

query is a python expression, using variables: socket, core, thread or "offline" to query all offline cpus format is a printf format with %d. %d will be replaced with the cpu number. format can be offline to offline the cpu or online to online

Print all cores on socket 0

cputop "socket == 0"

Print all first threads in each core on socket 0

cputop "thread == 0 and socket == 0"

Disable all second threads (disable hyper threading)

cputop "thread == 1" offline

Reenable all second threads (reenable hyper threading)

cputop "thread == 0" online

Self is a simple library to support self-profiling of programs, that is programs that measure their own execution.

cd self make

Read the documentation.

[rdpmc] (http://htmlpreview.github.com/?https://github.com/andikleen/pmu-tools/blob/master/self/rdpmc.html) is the basic facility to access raw counters.

ocperf can be used to generate raw perf numbers for your CPU to pass to rdpmc_open()

ocperf list | less

![example] (http://halobates.de/pmutools-event.png)

XXX is the needed event number in hex. Note that self does not support offcore or uncore events.

Also the event numbers are CPU specific, so you may need a /proc/cpuinfo model check for portable programs (see the ocperf source for example)

Example (replace EVENTNUMBER with your intended event from above or a perf event like PERF_COUNT_HW_CPU_CYCLES).

This is very simplified, for a real benchmark you almost certainly want some warmup, multiple iterations, possibly context switch filtering and some filler code to avoid cache effects.

	#include "rdpmc.h"

	struct rdpmc_ctx ctx;
	unsigned long long start, end;

	if (rdpmc_open(EVENTNUMBER, &ctx) < 0) ... error ...
	start = rdpmc_read(&ctx);
	... your workload ...
	end = rdpmc_read(&ctx);

[measure] (http://htmlpreview.github.com/?https://github.com/andikleen/pmu-tools/blob/master/self/measure.html) supports event group profiling.
[interrupts] (http://htmlpreview.github.com/?https://github.com/andikleen/pmu-tools/blob/master/self/interrupts.html) provides functions for a common use case of filtering out context switches and interrupts from micro benchmarks. These only work on Intel Ivy and Sandy Bridge CPUs.

Link the object files and include the header files in your program

/sys/devices/cpu/rdpmc must be 1.

rtest.c and test2.c provide examples. http://halobates.de/modern-pmus-yokohama.pdf provides some additional general information on cycle counting. The techniques used with simple-pmu described there can be used with self too.

Older perf doesn't export the raw PEBS output, which contains a lot of useful information. PEBS is a sampling format generated by Intel CPUs for some events.

pebs-grabber gabs PEBS data from perf. This assumes the perf pebs handler is running, we just also do trace points with the raw data.

May need some minor tweaks as kernel interface change, and will also not likely work on very old kernels.

This will create two new trace points trace_pebs_v1 and trace_pebs_v2 that log the complete PEBS record. When the CPU supports PEBSv2 (Haswell) the additional fields will be logged in pebs_v2.

  make [KDIR=/my/kernel/build/dir]
  insmod pebs-grabber.ko 
  # needs to record as root
  perf record -e cycles:p,pebs_v1,pebs\_v2 [command, -a for all etc.]
  perf report
  perf script to display pebs data
  # alternatively trace-cmd and kernelshark can be also used to dump
  # the pebs data

See http://download.intel.com/products/processor/manual/253669.pdf 18.10.2 for a description of the PEBS fields.

Note this doesn't work with standard FC18 kernels, as they broke trace points in modules. It works with later and earlier kernels.

addr is a basic frame work for self profiling of memory addresses accessed by the program. Requires a Linux 3.10+ kernel and a supported CPU.

Also provides a simple perf ring buffer API

event-rmap [cpu] prints the currently running events. This provides an easier answer to question Q2j in Vince Weaver's perf events FAQ.

Please post to the linux-perf-users@vger.kernel.org mailing list. For bugs please open an issue on https://github.com/andikleen/pmu-tools/issues

ocperf, toplev, ucevent are under GPLv2, self and addr under the modified BSD license.

Andi Kleen pmu-tools@halobates.de