xref: /linux/tools/perf/Documentation/topdown.txt (revision 891e8abed532423d3b918b0c445dc8919bc445b5)

Using TopDown metrics
---------------------

TopDown metrics break apart performance bottlenecks. Starting at level
1 it is typical to get metrics on retiring, bad speculation, frontend
bound, and backend bound. Higher levels provide more detail in to the
level 1 bottlenecks, such as at level 2: core bound, memory bound,
heavy operations, light operations, branch mispredicts, machine
clears, fetch latency and fetch bandwidth. For more details see [1][2][3].

perf stat --topdown implements this using available metrics that vary
per architecture.

% perf stat -a --topdown -I1000
#           time      %  tma_retiring %  tma_backend_bound %  tma_frontend_bound %  tma_bad_speculation
     1.001141351                 11.5                 34.9                  46.9                    6.7
     2.006141972                 13.4                 28.1                  50.4                    8.1
     3.010162040                 12.9                 28.1                  51.1                    8.0
     4.014009311                 12.5                 28.6                  51.8                    7.2
     5.017838554                 11.8                 33.0                  48.0                    7.2
     5.704818971                 14.0                 27.5                  51.3                    7.3
...

New Topdown features in Intel Ice Lake
======================================

With Ice Lake CPUs the TopDown metrics are directly available as
fixed counters and do not require generic counters. This allows
to collect TopDown always in addition to other events.

Using TopDown through RDPMC in applications on Intel Ice Lake
=============================================================

For more fine grained measurements it can be useful to
access the new  directly from user space. This is more complicated,
but drastically lowers overhead.

On Ice Lake, there is a new fixed counter 3: SLOTS, which reports
"pipeline SLOTS" (cycles multiplied by core issue width) and a
metric register that reports slots ratios for the different bottleneck
categories.

The metrics counter is CPU model specific and is not available on older
CPUs.

Example code
============

Library functions to do the functionality described below
is also available in libjevents [4]

The application opens a group with fixed counter 3 (SLOTS) and any
metric event, and allow user programs to read the performance counters.

Fixed counter 3 is mapped to a pseudo event event=0x00, umask=04,
so the perf_event_attr structure should be initialized with
{ .config = 0x0400, .type = PERF_TYPE_RAW }
The metric events are mapped to the pseudo event event=0x00, umask=0x8X.
For example, the perf_event_attr structure can be initialized with
{ .config = 0x8000, .type = PERF_TYPE_RAW } for Retiring metric event
The Fixed counter 3 must be the leader of the group.

#include <linux/perf_event.h>
#include <sys/mman.h>
#include <sys/syscall.h>
#include <unistd.h>

/* Provide own perf_event_open stub because glibc doesn't */
__attribute__((weak))
int perf_event_open(struct perf_event_attr *attr, pid_t pid,
		    int cpu, int group_fd, unsigned long flags)
{
	return syscall(__NR_perf_event_open, attr, pid, cpu, group_fd, flags);
}

/* Open slots counter file descriptor for current task. */
struct perf_event_attr slots = {
	.type = PERF_TYPE_RAW,
	.size = sizeof(struct perf_event_attr),
	.config = 0x400,
	.exclude_kernel = 1,
};

int slots_fd = perf_event_open(&slots, 0, -1, -1, 0);
if (slots_fd < 0)
	... error ...

/* Memory mapping the fd permits _rdpmc calls from userspace */
void *slots_p = mmap(0, getpagesize(), PROT_READ, MAP_SHARED, slots_fd, 0);
if (!slot_p)
	.... error ...

/*
 * Open metrics event file descriptor for current task.
 * Set slots event as the leader of the group.
 */
struct perf_event_attr metrics = {
	.type = PERF_TYPE_RAW,
	.size = sizeof(struct perf_event_attr),
	.config = 0x8000,
	.exclude_kernel = 1,
};

int metrics_fd = perf_event_open(&metrics, 0, -1, slots_fd, 0);
if (metrics_fd < 0)
	... error ...

/* Memory mapping the fd permits _rdpmc calls from userspace */
void *metrics_p = mmap(0, getpagesize(), PROT_READ, MAP_SHARED, metrics_fd, 0);
if (!metrics_p)
	... error ...

Note: the file descriptors returned by the perf_event_open calls must be memory
mapped to permit calls to the _rdpmd instruction. Permission may also be granted
by writing the /sys/devices/cpu/rdpmc sysfs node.

The RDPMC instruction (or _rdpmc compiler intrinsic) can now be used
to read slots and the topdown metrics at different points of the program:

#include <stdint.h>
#include <x86intrin.h>

#define RDPMC_FIXED	(1 << 30)	/* return fixed counters */
#define RDPMC_METRIC	(1 << 29)	/* return metric counters */

#define FIXED_COUNTER_SLOTS		3
#define METRIC_COUNTER_TOPDOWN_L1_L2	0

static inline uint64_t read_slots(void)
{
	return _rdpmc(RDPMC_FIXED | FIXED_COUNTER_SLOTS);
}

static inline uint64_t read_metrics(void)
{
	return _rdpmc(RDPMC_METRIC | METRIC_COUNTER_TOPDOWN_L1_L2);
}

Then the program can be instrumented to read these metrics at different
points.

It's not a good idea to do this with too short code regions,
as the parallelism and overlap in the CPU program execution will
cause too much measurement inaccuracy. For example instrumenting
individual basic blocks is definitely too fine grained.

_rdpmc calls should not be mixed with reading the metrics and slots counters
through system calls, as the kernel will reset these counters after each system
call.

Decoding metrics values
=======================

The value reported by read_metrics() contains four 8 bit fields
that represent a scaled ratio that represent the Level 1 bottleneck.
All four fields add up to 0xff (= 100%)

The binary ratios in the metric value can be converted to float ratios:

#define GET_METRIC(m, i) (((m) >> (i*8)) & 0xff)

/* L1 Topdown metric events */
#define TOPDOWN_RETIRING(val)	((float)GET_METRIC(val, 0) / 0xff)
#define TOPDOWN_BAD_SPEC(val)	((float)GET_METRIC(val, 1) / 0xff)
#define TOPDOWN_FE_BOUND(val)	((float)GET_METRIC(val, 2) / 0xff)
#define TOPDOWN_BE_BOUND(val)	((float)GET_METRIC(val, 3) / 0xff)

/*
 * L2 Topdown metric events.
 * Available on Sapphire Rapids and later platforms.
 */
#define TOPDOWN_HEAVY_OPS(val)		((float)GET_METRIC(val, 4) / 0xff)
#define TOPDOWN_BR_MISPREDICT(val)	((float)GET_METRIC(val, 5) / 0xff)
#define TOPDOWN_FETCH_LAT(val)		((float)GET_METRIC(val, 6) / 0xff)
#define TOPDOWN_MEM_BOUND(val)		((float)GET_METRIC(val, 7) / 0xff)

and then converted to percent for printing.

The ratios in the metric accumulate for the time when the counter
is enabled. For measuring programs it is often useful to measure
specific sections. For this it is needed to deltas on metrics.

This can be done by scaling the metrics with the slots counter
read at the same time.

Then it's possible to take deltas of these slots counts
measured at different points, and determine the metrics
for that time period.

	slots_a = read_slots();
	metric_a = read_metrics();

	... larger code region ...

	slots_b = read_slots()
	metric_b = read_metrics()

	# compute scaled metrics for measurement a
	retiring_slots_a = GET_METRIC(metric_a, 0) * slots_a
	bad_spec_slots_a = GET_METRIC(metric_a, 1) * slots_a
	fe_bound_slots_a = GET_METRIC(metric_a, 2) * slots_a
	be_bound_slots_a = GET_METRIC(metric_a, 3) * slots_a

	# compute delta scaled metrics between b and a
	retiring_slots = GET_METRIC(metric_b, 0) * slots_b - retiring_slots_a
	bad_spec_slots = GET_METRIC(metric_b, 1) * slots_b - bad_spec_slots_a
	fe_bound_slots = GET_METRIC(metric_b, 2) * slots_b - fe_bound_slots_a
	be_bound_slots = GET_METRIC(metric_b, 3) * slots_b - be_bound_slots_a

Later the individual ratios of L1 metric events for the measurement period can
be recreated from these counts.

	slots_delta = slots_b - slots_a
	retiring_ratio = (float)retiring_slots / slots_delta
	bad_spec_ratio = (float)bad_spec_slots / slots_delta
	fe_bound_ratio = (float)fe_bound_slots / slots_delta
	be_bound_ratio = (float)be_bound_slots / slota_delta

	printf("Retiring %.2f%% Bad Speculation %.2f%% FE Bound %.2f%% BE Bound %.2f%%\n",
		retiring_ratio * 100.,
		bad_spec_ratio * 100.,
		fe_bound_ratio * 100.,
		be_bound_ratio * 100.);

The individual ratios of L2 metric events for the measurement period can be
recreated from L1 and L2 metric counters. (Available on Sapphire Rapids and
later platforms)

	# compute scaled metrics for measurement a
	heavy_ops_slots_a = GET_METRIC(metric_a, 4) * slots_a
	br_mispredict_slots_a = GET_METRIC(metric_a, 5) * slots_a
	fetch_lat_slots_a = GET_METRIC(metric_a, 6) * slots_a
	mem_bound_slots_a = GET_METRIC(metric_a, 7) * slots_a

	# compute delta scaled metrics between b and a
	heavy_ops_slots = GET_METRIC(metric_b, 4) * slots_b - heavy_ops_slots_a
	br_mispredict_slots = GET_METRIC(metric_b, 5) * slots_b - br_mispredict_slots_a
	fetch_lat_slots = GET_METRIC(metric_b, 6) * slots_b - fetch_lat_slots_a
	mem_bound_slots = GET_METRIC(metric_b, 7) * slots_b - mem_bound_slots_a

	slots_delta = slots_b - slots_a
	heavy_ops_ratio = (float)heavy_ops_slots / slots_delta
	light_ops_ratio = retiring_ratio - heavy_ops_ratio;

	br_mispredict_ratio = (float)br_mispredict_slots / slots_delta
	machine_clears_ratio = bad_spec_ratio - br_mispredict_ratio;

	fetch_lat_ratio = (float)fetch_lat_slots / slots_delta
	fetch_bw_ratio = fe_bound_ratio - fetch_lat_ratio;

	mem_bound_ratio = (float)mem_bound_slots / slota_delta
	core_bound_ratio = be_bound_ratio - mem_bound_ratio;

	printf("Heavy Operations %.2f%% Light Operations %.2f%% "
	       "Branch Mispredict %.2f%% Machine Clears %.2f%% "
	       "Fetch Latency %.2f%% Fetch Bandwidth %.2f%% "
	       "Mem Bound %.2f%% Core Bound %.2f%%\n",
		heavy_ops_ratio * 100.,
		light_ops_ratio * 100.,
		br_mispredict_ratio * 100.,
		machine_clears_ratio * 100.,
		fetch_lat_ratio * 100.,
		fetch_bw_ratio * 100.,
		mem_bound_ratio * 100.,
		core_bound_ratio * 100.);

Resetting metrics counters
==========================

Since the individual metrics are only 8bit they lose precision for
short regions over time because the number of cycles covered by each
fraction bit shrinks. So the counters need to be reset regularly.

When using the kernel perf API the kernel resets on every read.
So as long as the reading is at reasonable intervals (every few
seconds) the precision is good.

When using perf stat it is recommended to always use the -I option,
with no longer interval than a few seconds

	perf stat -I 1000 --topdown ...

For user programs using RDPMC directly the counter can
be reset explicitly using ioctl:

	ioctl(perf_fd, PERF_EVENT_IOC_RESET, 0);

This "opens" a new measurement period.

A program using RDPMC for TopDown should schedule such a reset
regularly, as in every few seconds.

Limits on Intel Ice Lake
========================

Four pseudo TopDown metric events are exposed for the end-users,
topdown-retiring, topdown-bad-spec, topdown-fe-bound and topdown-be-bound.
They can be used to collect the TopDown value under the following
rules:
- All the TopDown metric events must be in a group with the SLOTS event.
- The SLOTS event must be the leader of the group.
- The PERF_FORMAT_GROUP flag must be applied for each TopDown metric
  events

The SLOTS event and the TopDown metric events can be counting members of
a sampling read group. Since the SLOTS event must be the leader of a TopDown
group, the second event of the group is the sampling event.
For example, perf record -e '{slots, $sampling_event, topdown-retiring}:S'

Extension on Intel Sapphire Rapids Server
=========================================
The metrics counter is extended to support TMA method level 2 metrics.
The lower half of the register is the TMA level 1 metrics (legacy).
The upper half is also divided into four 8-bit fields for the new level 2
metrics. Four more TopDown metric events are exposed for the end-users,
topdown-heavy-ops, topdown-br-mispredict, topdown-fetch-lat and
topdown-mem-bound.

Each of the new level 2 metrics in the upper half is a subset of the
corresponding level 1 metric in the lower half. Software can deduce the
other four level 2 metrics by subtracting corresponding metrics as below.

    Light_Operations = Retiring - Heavy_Operations
    Machine_Clears = Bad_Speculation - Branch_Mispredicts
    Fetch_Bandwidth = Frontend_Bound - Fetch_Latency
    Core_Bound = Backend_Bound - Memory_Bound

TPEBS in TopDown
================

TPEBS (Timed PEBS) is one of the new Intel PMU features provided since Granite
Rapids microarchitecture. The TPEBS feature adds a 16 bit retire_latency field
in the Basic Info group of the PEBS record. It records the Core cycles since the
retirement of the previous instruction to the retirement of current instruction.
Please refer to Section 8.4.1 of "Intel® Architecture Instruction Set Extensions
Programming Reference" for more details about this feature. Because this feature
extends PEBS record, sampling with weight option is required to get the
retire_latency value.

	perf record -e event_name -W ...

In the most recent release of TMA, the metrics begin to use event retire_latency
values in some of the metrics’ formulas on processors that support TPEBS feature.
For previous generations that do not support TPEBS, the values are static and
predefined per processor family by the hardware architects. Due to the diversity
of workloads in execution environments, retire_latency values measured at real
time are more accurate. Therefore, new TMA metrics that use TPEBS will provide
more accurate performance analysis results.

To support TPEBS in TMA metrics, a new modifier :R on event is added. Perf would
capture retire_latency value of required events(event with :R in metric formula)
with perf record. The retire_latency value would be used in metric calculation.
Currently, this feature is supported through perf stat

	perf stat -M metric_name --record-tpebs ...



[1] https://software.intel.com/en-us/top-down-microarchitecture-analysis-method-win
[2] https://sites.google.com/site/analysismethods/yasin-pubs
[3] https://perf.wiki.kernel.org/index.php/Top-Down_Analysis
[4] https://github.com/andikleen/pmu-tools/tree/master/jevents