[PATCH v4 1/6] x86,sched: Add support for frequency invariance

From: Giovanni Gherdovich
Date: Wed Nov 13 2019 - 07:41:24 EST

Implement arch_scale_freq_capacity() for 'modern' x86. This function
is used by the scheduler to correctly account usage in the face of
DVFS.

The present patch addresses Intel processors specifically and has positive
performance and performance-per-watt implications for the schedutil cpufreq
governor, bringing it closer to, if not on-par with, the powersave governor
from the intel_pstate driver/framework.

Large performance gains are obtained when the machine is lightly loaded and no
regression are observed at saturation. The benchmarks with the largest gains
are kernel compilation, tbench (the networking version of dbench) and
shell-intensive workloads.

1. FREQUENCY INVARIANCE: MOTIVATION
* Without it, a task looks larger if the CPU runs slower

2. PECULIARITIES OF X86
* freq invariance accounting requires knowing the ratio freq_curr/freq_max
2.1 CURRENT FREQUENCY
* Use delta_APERF / delta_MPERF * freq_base (a.k.a "BusyMHz")
2.2 MAX FREQUENCY
* It varies with time (turbo). As an approximation, we set it to a
constant, i.e. 4-cores turbo frequency (or base frequency if nothing
else is reported by MSRs)

3. EFFECTS ON THE SCHEDUTIL FREQUENCY GOVERNOR
* The invariant schedutil's formula has no feedback loop and reacts faster
to utilization changes

4. KNOWN LIMITATIONS
* In some cases tasks can't reach max util despite how hard they try

5. PERFORMANCE TESTING
5.1 MACHINES
* Skylake, Broadwell, Haswell
5.2 SETUP
* baseline Linux v5.2 w/ non-invariant schedutil. Tested freq_max = 1-2-3-4-8-12
active cores turbo w/ invariant schedutil, and intel_pstate/powersave
5.3 BENCHMARK RESULTS
5.3.1 NEUTRAL BENCHMARKS
* NAS Parallel Benchmark (HPC), hackbench
5.3.2 NON-NEUTRAL BENCHMARKS
* tbench (10-30% better), kernbench (10-15% better),
shell-intensive-scripts (30-50% better)
* no regressions
5.3.3 SELECTION OF DETAILED RESULTS
5.3.4 POWER CONSUMPTION, PERFORMANCE-PER-WATT
* dbench (5% worse on one machine), kernbench (3% worse),
tbench (5-10% better), shell-intensive-scripts (10-40% better)

6. MICROARCH'ES ADDRESSED HERE
* Xeon Core before Scalable Performance processors line (Xeon Gold/Platinum
etc have different MSRs semantic for querying turbo levels)

7. REFERENCES
* MMTests performance testing framework, github.com/gormanm/mmtests

+-------------------------------------------------------------------------+
| 1. FREQUENCY INVARIANCE: MOTIVATION
+-------------------------------------------------------------------------+

For example; suppose a CPU has two frequencies: 500 and 1000 Mhz. When
running a task that would consume 1/3rd of a CPU at 1000 MHz, it would
appear to consume 2/3rd (or 66.6%) when running at 500 MHz, giving the
false impression this CPU is almost at capacity, even though it can go
faster [*]. In a nutshell, without frequency scale-invariance tasks look
larger just because the CPU is running slower.

[*] (footnote: this assumes a linear frequency/performance relation; which
everybody knows to be false, but given realities its the best approximation
we can make.)

+-------------------------------------------------------------------------+
| 2. PECULIARITIES OF X86
+-------------------------------------------------------------------------+

Accounting for frequency changes in PELT signals requires the computation of
the ratio freq_curr / freq_max. On x86 neither of those terms is readily
available.

2.1 CURRENT FREQUENCY
====================

Since modern x86 has hardware control over the actual frequency we run
at (because amongst other things, Turbo-Mode), we cannot simply use
the frequency as requested through cpufreq.

Instead we use the APERF/MPERF MSRs to compute the effective frequency
over the recent past. Also, because reading MSRs is expensive, don't
do so every time we need the value, but amortize the cost by doing it
every tick.

2.2 MAX FREQUENCY
=================

Obtaining freq_max is also non-trivial because at any time the hardware can
provide a frequency boost to a selected subset of cores if the package has
enough power to spare (eg: Turbo Boost). This means that the maximum frequency
available to a given core changes with time.

The approach taken in this change is to arbitrarily set freq_max to a constant
value at boot. The value chosen is the "4-cores (4C) turbo frequency" on most
microarchitectures, after evaluating the following candidates:

* 1-core (1C) turbo frequency (the fastest turbo state available)
* around base frequency (a.k.a. max P-state)
* something in between, such as 4C turbo

To interpret these options, consider that this is the denominator in
freq_curr/freq_max, and that ratio will be used to scale PELT signals such as
util_avg and load_avg. A large denominator will undershoot (util_avg looks a
bit smaller than it really is), viceversa with a smaller denominator PELT
signals will tend to overshoot. Given that PELT drives frequency selection
in the schedutil governor, we will have:

freq_max set to | effect on DVFS
--------------------+------------------
1C turbo | power efficiency (lower freq choices)
base freq | performance (higher util_avg, higher freq requests)
4C turbo | a bit of both

4C turbo proves to be a good compromise in a number of benchmarks (see
below). Note that when the function intel_set_cpu_max_freq() fails to query
the various MSRs for the 4C turbo value, the variable arch_max_freq retains
its default value of SCHED_CAPACITY_SCALE (1024) that corresponds to setting
freq_max to base frequency wrt the table above.

+-------------------------------------------------------------------------+
| 3. EFFECTS ON THE SCHEDUTIL FREQUENCY GOVERNOR
+-------------------------------------------------------------------------+

Once an architecture implements a frequency scale-invariant utilization (the
PELT signal util_avg), schedutil switches its frequency selection formula from

freq_next = 1.25 * freq_curr * util [non-invariant util signal]

to

freq_next = 1.25 * freq_max * util [invariant util signal]

where, in the second formula, freq_max is set to the 1C turbo frequency (max
turbo). The advantage of the second formula, whose usage we unlock with this
patch, is that freq_next doesn't depend on the current frequency in an
iterative fashion, but can jump to any frequency in a single update. This
absence of feedback in the formula makes it quicker to react to utilization
changes and more robust against pathological instabilities.

Compare it to the update formula of intel_pstate/powersave:

freq_next = 1.25 * freq_max * Busy%

where again freq_max is 1C turbo and Busy% is the percentage of time not spent
idling (calculated with delta_MPERF / delta_TSC); essentially the same as
invariant schedutil, and largely responsible for intel_pstate/powersave good
reputation. The non-invariant schedutil formula is derived from the invariant
one by approximating util_inv with util_raw * freq_curr / freq_max, but this
has limitations.

Testing shows improved performances due to better frequency selections when
the machine is lightly loaded, and essentially no change in behaviour at
saturation / overutilization.

+-------------------------------------------------------------------------+
| 4. KNOWN LIMITATIONS
+-------------------------------------------------------------------------+

It's been shown that it is possible to create pathological scenarios where a
CPU-bound task cannot reach max utilization, if the normalizing factor
freq_max is fixed to a constant value (see [Lelli-2018]).

If freq_max is set to 4C turbo as we do here, one needs to peg at least 5
cores in a package doing some busywork, and observe that none of those task
will ever reach max util (1024) because they're all running at less than the
4C turbo frequency.

While this concern still applies, we believe the performance benefit of
frequency scale-invariant PELT signals outweights the cost of this limitation.

[Lelli-2018]
https://lore.kernel.org/lkml/20180517150418.GF22493@xxxxxxxxxxxxxxxxxxxxx/

+-------------------------------------------------------------------------+
| 5. PERFORMANCE TESTING
+-------------------------------------------------------------------------+

5.1 MACHINES
============

We tested the patch on three machines, with Skylake, Broadwell and Haswell
CPUs. The details are below, together with the available turbo ratios as
reported by the appropriate MSRs.

* 8x-SKYLAKE-UMA:
Single socket E3-1240 v5, Skylake 4 cores/8 threads
Max EFFiciency, BASE frequency and available turbo levels (MHz):

EFFIC 800 |********
BASE 3500 |***********************************
4C 3700 |*************************************
3C 3800 |**************************************
2C 3900 |***************************************
1C 3900 |***************************************

* 80x-BROADWELL-NUMA:
Two sockets E5-2698 v4, 2x Broadwell 20 cores/40 threads
Max EFFiciency, BASE frequency and available turbo levels (MHz):

EFFIC 1200 |************
BASE 2200 |**********************
8C 2900 |*****************************
7C 3000 |******************************
6C 3100 |*******************************
5C 3200 |********************************
4C 3300 |*********************************
3C 3400 |**********************************
2C 3600 |************************************
1C 3600 |************************************

* 48x-HASWELL-NUMA
Two sockets E5-2670 v3, 2x Haswell 12 cores/24 threads
Max EFFiciency, BASE frequency and available turbo levels (MHz):

EFFIC 1200 |************
BASE 2300 |***********************
12C 2600 |**************************
11C 2600 |**************************
10C 2600 |**************************
9C 2600 |**************************
8C 2600 |**************************
7C 2600 |**************************
6C 2600 |**************************
5C 2700 |***************************
4C 2800 |****************************
3C 2900 |*****************************
2C 3100 |*******************************
1C 3100 |*******************************

5.2 SETUP
=========

* The baseline is Linux v5.2 with schedutil (non-invariant) and the intel_pstate
driver in passive mode.
* The rationale for choosing the various freq_max values to test have been to
try all the 1-2-3-4C turbo levels (note that 1C and 2C turbo are identical
on all machines), plus one more value closer to base_freq but still in the
turbo range (8C turbo for both 80x-BROADWELL-NUMA and 48x-HASWELL-NUMA).
* In addition we've run all tests with intel_pstate/powersave for comparison.
* The filesystem is always XFS, the userspace is openSUSE Leap 15.1.
* 8x-SKYLAKE-UMA is capable of HWP (Hardware-Managed P-States), so the runs
with active intel_pstate on this machine use that.

This gives, in terms of combinations tested on each machine:

* 8x-SKYLAKE-UMA
* Baseline: Linux v5.2, non-invariant schedutil, intel_pstate passive
* intel_pstate active + powersave + HWP
* invariant schedutil, freq_max = 1C turbo
* invariant schedutil, freq_max = 3C turbo
* invariant schedutil, freq_max = 4C turbo

* both 80x-BROADWELL-NUMA and 48x-HASWELL-NUMA
* [same as 8x-SKYLAKE-UMA, but no HWP capable]
* invariant schedutil, freq_max = 8C turbo
* (which on 48x-HASWELL-NUMA is the same as 12C turbo, or "all cores turbo")

5.3 BENCHMARK RESULTS
=====================

5.3.1 NEUTRAL BENCHMARKS
------------------------

Tests that didn't show any measurable difference in performance on any of the
test machines between non-invariant schedutil and our patch are:

* NAS Parallel Benchmarks (NPB) using either MPI or openMP for IPC, any
computational kernel
* flexible I/O (FIO)
* hackbench (using threads or processes, and using pipes or sockets)

5.3.2 NON-NEUTRAL BENCHMARKS
----------------------------

What follow are summary tables where each benchmark result is given a score.

* A tilde (~) means a neutral result, i.e. no difference from baseline.
* Scores are computed with the ratio result_new / result_baseline, so a tilde
means a score of 1.00.
* The results in the score ratio are the geometric means of results running
the benchmark with different parameters (eg: for kernbench: using 1, 2, 4,
... number of processes; for pgbench: varying the number of clients, and so
on).
* The first three tables show higher-is-better kind of tests (i.e. measured in
operations/second), the subsequent three show lower-is-better kind of tests
(i.e. the workload is fixed and we measure elapsed time, think kernbench).
* "gitsource" is a name we made up for the test consisting in running the
entire unit tests suite of the Git SCM and measuring how long it takes. We
take it as a typical example of shell-intensive serialized workload.
* In the "I_PSTATE" column we have the results for intel_pstate/powersave. Other
columns show invariant schedutil for different values of freq_max. 4C turbo
is circled as it's the value we've chosen for the final implementation.

80x-BROADWELL-NUMA (comparison ratio; higher is better)
+------+
I_PSTATE 1C 3C | 4C | 8C
pgbench-ro 1.14 ~ ~ | 1.11 | 1.14
pgbench-rw ~ ~ ~ | ~ | ~
netperf-udp 1.06 ~ 1.06 | 1.05 | 1.07
netperf-tcp ~ 1.03 ~ | 1.01 | 1.02
tbench4 1.57 1.18 1.22 | 1.30 | 1.56
+------+

8x-SKYLAKE-UMA (comparison ratio; higher is better)
+------+
I_PSTATE/HWP 1C 3C | 4C |
pgbench-ro ~ ~ ~ | ~ |
pgbench-rw ~ ~ ~ | ~ |
netperf-udp ~ ~ ~ | ~ |
netperf-tcp ~ ~ ~ | ~ |
tbench4 1.30 1.14 1.14 | 1.16 |
+------+

48x-HASWELL-NUMA (comparison ratio; higher is better)
+------+
I_PSTATE 1C 3C | 4C | 12C
pgbench-ro 1.15 ~ ~ | 1.06 | 1.16
pgbench-rw ~ ~ ~ | ~ | ~
netperf-udp 1.05 0.97 1.04 | 1.04 | 1.02
netperf-tcp 0.96 1.01 1.01 | 1.01 | 1.01
tbench4 1.50 1.05 1.13 | 1.13 | 1.25
+------+

In the table above we see that active intel_pstate is slightly better than our
4C-turbo patch (both in reference to the baseline non-invariant schedutil) on
read-only pgbench and much better on tbench. Both cases are notable in which
it shows that lowering our freq_max (to 8C-turbo and 12C-turbo on
80x-BROADWELL-NUMA and 48x-HASWELL-NUMA respectively) helps invariant
schedutil to get closer.

If we ignore active intel_pstate and focus on the comparison with baseline
alone, there are several instances of double-digit performance improvement.

80x-BROADWELL-NUMA (comparison ratio; lower is better)
+------+
I_PSTATE 1C 3C | 4C | 8C
dbench4 1.23 0.95 0.95 | 0.95 | 0.95
kernbench 0.93 0.83 0.83 | 0.83 | 0.82
gitsource 0.98 0.49 0.49 | 0.49 | 0.48
+------+

8x-SKYLAKE-UMA (comparison ratio; lower is better)
+------+
I_PSTATE/HWP 1C 3C | 4C |
dbench4 ~ ~ ~ | ~ |
kernbench ~ ~ ~ | ~ |
gitsource 0.92 0.55 0.55 | 0.55 |
+------+

48x-HASWELL-NUMA (comparison ratio; lower is better)
+------+
I_PSTATE 1C 3C | 4C | 8C
dbench4 ~ ~ ~ | ~ | ~
kernbench 0.94 0.90 0.89 | 0.90 | 0.90
gitsource 0.97 0.69 0.69 | 0.69 | 0.69
+------+

dbench is not very remarkable here, unless we notice how poorly active
intel_pstate is performing on 80x-BROADWELL-NUMA: 23% regression versus
non-invariant schedutil. We repeated that run getting consistent results. Out
of scope for the patch at hand, but deserving future investigation. Other than
that, we previously ran this campaign with Linux v5.0 and saw the patch doing
better on dbench a the time. We haven't checked closely and can only speculate
at this point.

On the NUMA boxes kernbench gets 10-15% improvements on average; we'll see in
the detailed tables that the gains concentrate on low process counts (lightly
loaded machines).

The test we call "gitsource" (running the git unit test suite, a long-running
single-threaded shell script) appears rather spectacular in this table (gains
of 30-50% depending on the machine). It is to be noted, however, that
gitsource has no adjustable parameters (such as the number of jobs in
kernbench, which we average over in order to get a single-number summary
score) and is exactly the kind of low-parallelism workload that benefits the
most from this patch. When looking at the detailed tables of kernbench or
tbench4, at low process or client counts one can see similar numbers.

5.3.3 SELECTION OF DETAILED RESULTS
-----------------------------------

Machine : 48x-HASWELL-NUMA
Benchmark : tbench4 (i.e. dbench4 over the network, actually loopback)
Varying parameter : number of clients
Unit : MB/sec (higher is better)

5.2.0 vanilla (BASELINE) 5.2.0 intel_pstate 5.2.0 1C-turbo
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Hmean 1 126.73 +- 0.31% ( ) 315.91 +- 0.66% ( 149.28%) 125.03 +- 0.76% ( -1.34%)
Hmean 2 258.04 +- 0.62% ( ) 614.16 +- 0.51% ( 138.01%) 269.58 +- 1.45% ( 4.47%)
Hmean 4 514.30 +- 0.67% ( ) 1146.58 +- 0.54% ( 122.94%) 533.84 +- 1.99% ( 3.80%)
Hmean 8 1111.38 +- 2.52% ( ) 2159.78 +- 0.38% ( 94.33%) 1359.92 +- 1.56% ( 22.36%)
Hmean 16 2286.47 +- 1.36% ( ) 3338.29 +- 0.21% ( 46.00%) 2720.20 +- 0.52% ( 18.97%)
Hmean 32 4704.84 +- 0.35% ( ) 4759.03 +- 0.43% ( 1.15%) 4774.48 +- 0.30% ( 1.48%)
Hmean 64 7578.04 +- 0.27% ( ) 7533.70 +- 0.43% ( -0.59%) 7462.17 +- 0.65% ( -1.53%)
Hmean 128 6998.52 +- 0.16% ( ) 6987.59 +- 0.12% ( -0.16%) 6909.17 +- 0.14% ( -1.28%)
Hmean 192 6901.35 +- 0.25% ( ) 6913.16 +- 0.10% ( 0.17%) 6855.47 +- 0.21% ( -0.66%)

5.2.0 3C-turbo 5.2.0 4C-turbo 5.2.0 12C-turbo
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Hmean 1 128.43 +- 0.28% ( 1.34%) 130.64 +- 3.81% ( 3.09%) 153.71 +- 5.89% ( 21.30%)
Hmean 2 311.70 +- 6.15% ( 20.79%) 281.66 +- 3.40% ( 9.15%) 305.08 +- 5.70% ( 18.23%)
Hmean 4 641.98 +- 2.32% ( 24.83%) 623.88 +- 5.28% ( 21.31%) 906.84 +- 4.65% ( 76.32%)
Hmean 8 1633.31 +- 1.56% ( 46.96%) 1714.16 +- 0.93% ( 54.24%) 2095.74 +- 0.47% ( 88.57%)
Hmean 16 3047.24 +- 0.42% ( 33.27%) 3155.02 +- 0.30% ( 37.99%) 3634.58 +- 0.15% ( 58.96%)
Hmean 32 4734.31 +- 0.60% ( 0.63%) 4804.38 +- 0.23% ( 2.12%) 4674.62 +- 0.27% ( -0.64%)
Hmean 64 7699.74 +- 0.35% ( 1.61%) 7499.72 +- 0.34% ( -1.03%) 7659.03 +- 0.25% ( 1.07%)
Hmean 128 6935.18 +- 0.15% ( -0.91%) 6942.54 +- 0.10% ( -0.80%) 7004.85 +- 0.12% ( 0.09%)
Hmean 192 6901.62 +- 0.12% ( 0.00%) 6856.93 +- 0.10% ( -0.64%) 6978.74 +- 0.10% ( 1.12%)

This is one of the cases where the patch still can't surpass active
intel_pstate, not even when freq_max is as low as 12C-turbo. Otherwise, gains are
visible up to 16 clients and the saturated scenario is the same as baseline.

The scores in the summary table from the previous sections are ratios of
geometric means of the results over different clients, as seen in this table.

Machine : 80x-BROADWELL-NUMA
Benchmark : kernbench (kernel compilation)
Varying parameter : number of jobs
Unit : seconds (lower is better)

5.2.0 vanilla (BASELINE) 5.2.0 intel_pstate 5.2.0 1C-turbo
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Amean 2 379.68 +- 0.06% ( ) 330.20 +- 0.43% ( 13.03%) 285.93 +- 0.07% ( 24.69%)
Amean 4 200.15 +- 0.24% ( ) 175.89 +- 0.22% ( 12.12%) 153.78 +- 0.25% ( 23.17%)
Amean 8 106.20 +- 0.31% ( ) 95.54 +- 0.23% ( 10.03%) 86.74 +- 0.10% ( 18.32%)
Amean 16 56.96 +- 1.31% ( ) 53.25 +- 1.22% ( 6.50%) 48.34 +- 1.73% ( 15.13%)
Amean 32 34.80 +- 2.46% ( ) 33.81 +- 0.77% ( 2.83%) 30.28 +- 1.59% ( 12.99%)
Amean 64 26.11 +- 1.63% ( ) 25.04 +- 1.07% ( 4.10%) 22.41 +- 2.37% ( 14.16%)
Amean 128 24.80 +- 1.36% ( ) 23.57 +- 1.23% ( 4.93%) 21.44 +- 1.37% ( 13.55%)
Amean 160 24.85 +- 0.56% ( ) 23.85 +- 1.17% ( 4.06%) 21.25 +- 1.12% ( 14.49%)

5.2.0 3C-turbo 5.2.0 4C-turbo 5.2.0 8C-turbo
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Amean 2 284.08 +- 0.13% ( 25.18%) 283.96 +- 0.51% ( 25.21%) 285.05 +- 0.21% ( 24.92%)
Amean 4 153.18 +- 0.22% ( 23.47%) 154.70 +- 1.64% ( 22.71%) 153.64 +- 0.30% ( 23.24%)
Amean 8 87.06 +- 0.28% ( 18.02%) 86.77 +- 0.46% ( 18.29%) 86.78 +- 0.22% ( 18.28%)
Amean 16 48.03 +- 0.93% ( 15.68%) 47.75 +- 1.99% ( 16.17%) 47.52 +- 1.61% ( 16.57%)
Amean 32 30.23 +- 1.20% ( 13.14%) 30.08 +- 1.67% ( 13.57%) 30.07 +- 1.67% ( 13.60%)
Amean 64 22.59 +- 2.02% ( 13.50%) 22.63 +- 0.81% ( 13.32%) 22.42 +- 0.76% ( 14.12%)
Amean 128 21.37 +- 0.67% ( 13.82%) 21.31 +- 1.15% ( 14.07%) 21.17 +- 1.93% ( 14.63%)
Amean 160 21.68 +- 0.57% ( 12.76%) 21.18 +- 1.74% ( 14.77%) 21.22 +- 1.00% ( 14.61%)

The patch outperform active intel_pstate (and baseline) by a considerable
margin; the summary table from the previous section says 4C turbo and active
intel_pstate are 0.83 and 0.93 against baseline respectively, so 4C turbo is
0.83/0.93=0.89 against intel_pstate (~10% better on average). There is no
noticeable difference with regard to the value of freq_max.

Machine : 8x-SKYLAKE-UMA
Benchmark : gitsource (time to run the git unit test suite)
Varying parameter : none
Unit : seconds (lower is better)

5.2.0 vanilla 5.2.0 intel_pstate/hwp 5.2.0 1C-turbo
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Amean 858.85 +- 1.16% ( ) 791.94 +- 0.21% ( 7.79%) 474.95 ( 44.70%)

5.2.0 3C-turbo 5.2.0 4C-turbo
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Amean 475.26 +- 0.20% ( 44.66%) 474.34 +- 0.13% ( 44.77%)

In this test, which is of interest as representing shell-intensive
(i.e. fork-intensive) serialized workloads, invariant schedutil outperforms
intel_pstate/powersave by a whopping 40% margin.

5.3.4 POWER CONSUMPTION, PERFORMANCE-PER-WATT
---------------------------------------------

The following table shows average power consumption in watt for each
benchmark. Data comes from turbostat (package average), which in turn is read
from the RAPL interface on CPUs. We know the patch affects CPU frequencies so
it's reasonable to ignore other power consumers (such as memory or I/O). Also,
we don't have a power meter available in the lab so RAPL is the best we have.

turbostat sampled average power every 10 seconds for the entire duration of
each benchmark. We took all those values and averaged them (i.e. with don't
have detail on a per-parameter granularity, only on whole benchmarks).

80x-BROADWELL-NUMA (power consumption, watts)
+--------+
BASELINE I_PSTATE 1C 3C | 4C | 8C
pgbench-ro 130.01 142.77 131.11 132.45 | 134.65 | 136.84
pgbench-rw 68.30 60.83 71.45 71.70 | 71.65 | 72.54
dbench4 90.25 59.06 101.43 99.89 | 101.10 | 102.94
netperf-udp 65.70 69.81 66.02 68.03 | 68.27 | 68.95
netperf-tcp 88.08 87.96 88.97 88.89 | 88.85 | 88.20
tbench4 142.32 176.73 153.02 163.91 | 165.58 | 176.07
kernbench 92.94 101.95 114.91 115.47 | 115.52 | 115.10
gitsource 40.92 41.87 75.14 75.20 | 75.40 | 75.70
+--------+
8x-SKYLAKE-UMA (power consumption, watts)
+--------+
BASELINE I_PSTATE/HWP 1C 3C | 4C |
pgbench-ro 46.49 46.68 46.56 46.59 | 46.52 |
pgbench-rw 29.34 31.38 30.98 31.00 | 31.00 |
dbench4 27.28 27.37 27.49 27.41 | 27.38 |
netperf-udp 22.33 22.41 22.36 22.35 | 22.36 |
netperf-tcp 27.29 27.29 27.30 27.31 | 27.33 |
tbench4 41.13 45.61 43.10 43.33 | 43.56 |
kernbench 42.56 42.63 43.01 43.01 | 43.01 |
gitsource 13.32 13.69 17.33 17.30 | 17.35 |
+--------+
48x-HASWELL-NUMA (power consumption, watts)
+--------+
BASELINE I_PSTATE 1C 3C | 4C | 12C
pgbench-ro 128.84 136.04 129.87 132.43 | 132.30 | 134.86
pgbench-rw 37.68 37.92 37.17 37.74 | 37.73 | 37.31
dbench4 28.56 28.73 28.60 28.73 | 28.70 | 28.79
netperf-udp 56.70 60.44 56.79 57.42 | 57.54 | 57.52
netperf-tcp 75.49 75.27 75.87 76.02 | 76.01 | 75.95
tbench4 115.44 139.51 119.53 123.07 | 123.97 | 130.22
kernbench 83.23 91.55 95.58 95.69 | 95.72 | 96.04
gitsource 36.79 36.99 39.99 40.34 | 40.35 | 40.23
+--------+

A lower power consumption isn't necessarily better, it depends on what is done
with that energy. Here are tables with the ratio of performance-per-watt on
each machine and benchmark. Higher is always better; a tilde (~) means a
neutral ratio (i.e. 1.00).

80x-BROADWELL-NUMA (performance-per-watt ratios; higher is better)
+------+
I_PSTATE 1C 3C | 4C | 8C
pgbench-ro 1.04 1.06 0.94 | 1.07 | 1.08
pgbench-rw 1.10 0.97 0.96 | 0.96 | 0.97
dbench4 1.24 0.94 0.95 | 0.94 | 0.92
netperf-udp ~ 1.02 1.02 | ~ | 1.02
netperf-tcp ~ 1.02 ~ | ~ | 1.02
tbench4 1.26 1.10 1.06 | 1.12 | 1.26
kernbench 0.98 0.97 0.97 | 0.97 | 0.98
gitsource ~ 1.11 1.11 | 1.11 | 1.13
+------+

8x-SKYLAKE-UMA (performance-per-watt ratios; higher is better)
+------+
I_PSTATE/HWP 1C 3C | 4C |
pgbench-ro ~ ~ ~ | ~ |
pgbench-rw 0.95 0.97 0.96 | 0.96 |
dbench4 ~ ~ ~ | ~ |
netperf-udp ~ ~ ~ | ~ |
netperf-tcp ~ ~ ~ | ~ |
tbench4 1.17 1.09 1.08 | 1.10 |
kernbench ~ ~ ~ | ~ |
gitsource 1.06 1.40 1.40 | 1.40 |
+------+

48x-HASWELL-NUMA (performance-per-watt ratios; higher is better)
+------+
I_PSTATE 1C 3C | 4C | 12C
pgbench-ro 1.09 ~ 1.09 | 1.03 | 1.11
pgbench-rw ~ 0.86 ~ | ~ | 0.86
dbench4 ~ 1.02 1.02 | 1.02 | ~
netperf-udp ~ 0.97 1.03 | 1.02 | ~
netperf-tcp 0.96 ~ ~ | ~ | ~
tbench4 1.24 ~ 1.06 | 1.05 | 1.11
kernbench 0.97 0.97 0.98 | 0.97 | 0.96
gitsource 1.03 1.33 1.32 | 1.32 | 1.33
+------+

These results are overall pleasing: in plenty of cases we observe
performance-per-watt improvements. The few regressions (read/write pgbench and
dbench on the Broadwell machine) are of small magnitude. kernbench loses a few
percentage points (it has a 10-15% performance improvement, but apparently the
increase in power consumption is larger than that). tbench4 and gitsource, which
benefit the most from the patch, keep a positive score in this table which is
a welcome surprise; that suggests that in those particular workloads the
non-invariant schedutil (and active intel_pstate, too) makes some rather
suboptimal frequency selections.

+-------------------------------------------------------------------------+
| 6. MICROARCH'ES ADDRESSED HERE
+-------------------------------------------------------------------------+

The patch addresses Xeon Core processors that use MSR_PLATFORM_INFO and
MSR_TURBO_RATIO_LIMIT to advertise their base frequency and turbo frequencies
respectively. This excludes the recent Xeon Scalable Performance processors
line (Xeon Gold, Platinum etc) whose MSRs have to be parsed differently.

Subsequent patches will address:

* Xeon Scalable Performance processors and Atom Goldmont/Goldmont Plus
* Xeon Phi (Knights Landing, Knights Mill)
* Atom Silvermont

+-------------------------------------------------------------------------+
| 7. REFERENCES
+-------------------------------------------------------------------------+

Tests have been run with the help of the MMTests performance testing
framework, see github.com/gormanm/mmtests. The configuration file names for
the benchmark used are:

db-pgbench-timed-ro-small-xfs
db-pgbench-timed-rw-small-xfs
io-dbench4-async-xfs
network-netperf-unbound
network-tbench
scheduler-unbound
workload-kerndevel-xfs
workload-shellscripts-xfs
hpc-nas-c-class-mpi-full-xfs
hpc-nas-c-class-omp-full

All those benchmarks are generally available on the web:

pgbench: https://www.postgresql.org/docs/10/pgbench.html
netperf: https://hewlettpackard.github.io/netperf/
dbench/tbench: https://dbench.samba.org/
gitsource: git unit test suite, github.com/git/git
NAS Parallel Benchmarks: https://www.nas.nasa.gov/publications/npb.html
hackbench: https://people.redhat.com/mingo/cfs-scheduler/tools/hackbench.c

Suggested-by: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Signed-off-by: Giovanni Gherdovich <ggherdovich@xxxxxxx>
Acked-by: Doug Smythies <dsmythies@xxxxxxxxx>
---
arch/x86/include/asm/topology.h | 23 ++++++
arch/x86/kernel/smpboot.c | 176 +++++++++++++++++++++++++++++++++++++++-
kernel/sched/core.c | 1 +
kernel/sched/sched.h | 7 ++
4 files changed, 206 insertions(+), 1 deletion(-)

diff --git a/arch/x86/include/asm/topology.h b/arch/x86/include/asm/topology.h
index 4b14d2318251..9b3aca463c8f 100644
--- a/arch/x86/include/asm/topology.h
+++ b/arch/x86/include/asm/topology.h
@@ -193,4 +193,27 @@ static inline void sched_clear_itmt_support(void)
}
#endif /* CONFIG_SCHED_MC_PRIO */

+#ifdef CONFIG_SMP
+#include <asm/cpufeature.h>
+
+DECLARE_STATIC_KEY_FALSE(arch_scale_freq_key);
+
+#define arch_scale_freq_invariant() static_branch_likely(&arch_scale_freq_key)
+
+DECLARE_PER_CPU(unsigned long, arch_cpu_freq);
+
+static inline long arch_scale_freq_capacity(int cpu)
+{
+ if (arch_scale_freq_invariant())
+ return per_cpu(arch_cpu_freq, cpu);
+
+ return 1024 /* SCHED_CAPACITY_SCALE */;
+}
+#define arch_scale_freq_capacity arch_scale_freq_capacity
+
+extern void arch_scale_freq_tick(void);
+#define arch_scale_freq_tick arch_scale_freq_tick
+
+#endif
+
#endif /* _ASM_X86_TOPOLOGY_H */
diff --git a/arch/x86/kernel/smpboot.c b/arch/x86/kernel/smpboot.c
index 69881b2d446c..814d7900779d 100644
--- a/arch/x86/kernel/smpboot.c
+++ b/arch/x86/kernel/smpboot.c
@@ -147,6 +147,8 @@ static inline void smpboot_restore_warm_reset_vector(void)
*((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
}

+static void set_cpu_max_freq(void);
+
/*
* Report back to the Boot Processor during boot time or to the caller processor
* during CPU online.
@@ -183,6 +185,8 @@ static void smp_callin(void)
*/
set_cpu_sibling_map(raw_smp_processor_id());

+ set_cpu_max_freq();
+
/*
* Get our bogomips.
* Update loops_per_jiffy in cpu_data. Previous call to
@@ -1337,7 +1341,7 @@ void __init native_smp_prepare_cpus(unsigned int max_cpus)
set_sched_topology(x86_topology);

set_cpu_sibling_map(0);
-
+ set_cpu_max_freq();
smp_sanity_check();

switch (apic_intr_mode) {
@@ -1764,3 +1768,173 @@ void native_play_dead(void)
}

#endif
+
+/*
+ * APERF/MPERF frequency ratio computation.
+ *
+ * The scheduler wants to do frequency invariant accounting and needs a <1
+ * ratio to account for the 'current' frequency, corresponding to
+ * freq_curr / freq_max.
+ *
+ * Since the frequency freq_curr on x86 is controlled by micro-controller and
+ * our P-state setting is little more than a request/hint, we need to observe
+ * the effective frequency 'BusyMHz', i.e. the average frequency over a time
+ * interval after discarding idle time. This is given by:
+ *
+ * BusyMHz = delta_APERF / delta_MPERF * freq_base
+ *
+ * where freq_base is the max non-turbo P-state.
+ *
+ * The freq_max term has to be set to a somewhat arbitrary value, because we
+ * can't know which turbo states will be available at a given point in time:
+ * it all depends on the thermal headroom of the entire package. We set it to
+ * the turbo level with 4 cores active.
+ *
+ * Benchmarks show that's a good compromise between the 1C turbo ratio
+ * (freq_curr/freq_max would rarely reach 1) and something close to freq_base,
+ * which would ignore the entire turbo range (a conspicuous part, making
+ * freq_curr/freq_max always maxed out).
+ *
+ * Setting freq_max to anything less than the 1C turbo ratio makes the ratio
+ * freq_curr / freq_max to eventually grow >1, in which case we clip it to 1.
+ */
+
+DEFINE_STATIC_KEY_FALSE(arch_scale_freq_key);
+
+static DEFINE_PER_CPU(u64, arch_prev_aperf);
+static DEFINE_PER_CPU(u64, arch_prev_mperf);
+static u64 arch_max_freq = SCHED_CAPACITY_SCALE;
+
+static bool turbo_disabled(void)
+{
+ u64 misc_en;
+ int err;
+
+ err = rdmsrl_safe(MSR_IA32_MISC_ENABLE, &misc_en);
+ if (err)
+ return false;
+
+ return (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
+}
+
+#include <asm/cpu_device_id.h>
+#include <asm/intel-family.h>
+
+#define ICPU(model) \
+ {X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF, 0}
+
+static const struct x86_cpu_id has_knl_turbo_ratio_limits[] = {
+ ICPU(INTEL_FAM6_XEON_PHI_KNL),
+ ICPU(INTEL_FAM6_XEON_PHI_KNM),
+ {}
+};
+
+static const struct x86_cpu_id has_skx_turbo_ratio_limits[] = {
+ ICPU(INTEL_FAM6_SKYLAKE_X),
+ {}
+};
+
+static const struct x86_cpu_id has_glm_turbo_ratio_limits[] = {
+ ICPU(INTEL_FAM6_ATOM_GOLDMONT),
+ ICPU(INTEL_FAM6_ATOM_GOLDMONT_D),
+ ICPU(INTEL_FAM6_ATOM_GOLDMONT_PLUS),
+ {}
+};
+
+static void core_set_cpu_max_freq(void)
+{
+ u64 ratio, turbo_ratio;
+ int err;
+
+ err = rdmsrl_safe(MSR_PLATFORM_INFO, &ratio);
+ if (err)
+ return;
+
+ err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT, &turbo_ratio);
+ if (err)
+ return;
+
+ ratio = (ratio >> 8) & 0xFF; /* max P state ratio */
+ turbo_ratio = (turbo_ratio >> 24) & 0xFF; /* 4C turbo ratio */
+
+ arch_max_freq = div_u64(turbo_ratio * SCHED_CAPACITY_SCALE, ratio);
+}
+
+static void intel_set_cpu_max_freq(void)
+{
+ /*
+ * TODO: add support for:
+ *
+ * - Xeon Gold/Platinum
+ * - Xeon Phi (KNM, KNL)
+ * - Atom Goldmont
+ * - Atom Silvermont
+ *
+ * which all now get by default arch_max_freq = SCHED_CAPACITY_SCALE
+ */
+
+ static_branch_enable(&arch_scale_freq_key);
+
+ if (turbo_disabled() ||
+ x86_match_cpu(has_skx_turbo_ratio_limits) ||
+ x86_match_cpu(has_knl_turbo_ratio_limits) ||
+ x86_match_cpu(has_glm_turbo_ratio_limits))
+ return;
+
+ core_set_cpu_max_freq();
+}
+
+static void init_scale_freq(void *arg)
+{
+ u64 aperf, mperf;
+
+ rdmsrl(MSR_IA32_APERF, aperf);
+ rdmsrl(MSR_IA32_MPERF, mperf);
+
+ this_cpu_write(arch_prev_aperf, aperf);
+ this_cpu_write(arch_prev_mperf, mperf);
+}
+
+static void set_cpu_max_freq(void)
+{
+ if (smp_processor_id() != 0 || !boot_cpu_has(X86_FEATURE_APERFMPERF))
+ return;
+
+ if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
+ intel_set_cpu_max_freq();
+
+ init_scale_freq(NULL);
+}
+
+DEFINE_PER_CPU(unsigned long, arch_cpu_freq);
+
+void arch_scale_freq_tick(void)
+{
+ u64 freq;
+ u64 aperf, mperf;
+ u64 acnt, mcnt;
+
+ if (!arch_scale_freq_invariant())
+ return;
+
+ rdmsrl(MSR_IA32_APERF, aperf);
+ rdmsrl(MSR_IA32_MPERF, mperf);
+
+ acnt = aperf - this_cpu_read(arch_prev_aperf);
+ mcnt = mperf - this_cpu_read(arch_prev_mperf);
+ if (!mcnt)
+ return;
+
+ this_cpu_write(arch_prev_aperf, aperf);
+ this_cpu_write(arch_prev_mperf, mperf);
+
+ acnt <<= 2*SCHED_CAPACITY_SHIFT;
+ mcnt *= arch_max_freq;
+
+ freq = div64_u64(acnt, mcnt);
+
+ if (freq > SCHED_CAPACITY_SCALE)
+ freq = SCHED_CAPACITY_SCALE;
+
+ this_cpu_write(arch_cpu_freq, freq);
+}
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 0f2eb3629070..1d0f5df8020e 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -3593,6 +3593,7 @@ void scheduler_tick(void)
struct task_struct *curr = rq->curr;
struct rq_flags rf;

+ arch_scale_freq_tick();
sched_clock_tick();

rq_lock(rq, &rf);
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index c8870c5bd7df..4c4f5f197e10 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -1977,6 +1977,13 @@ static inline int hrtick_enabled(struct rq *rq)

#endif /* CONFIG_SCHED_HRTICK */

+#ifndef arch_scale_freq_tick
+static __always_inline
+void arch_scale_freq_tick(void)
+{
+}
+#endif
+
#ifndef arch_scale_freq_capacity
static __always_inline
unsigned long arch_scale_freq_capacity(int cpu)
--
2.16.4