Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
From:   Shijith Thotton <sthotton@marvell.com>
To:     Steve Sistare <steven.sistare@oracle.com>,
        "mingo@redhat.com" <mingo@redhat.com>,
        "peterz@infradead.org" <peterz@infradead.org>
CC:     "subhra.mazumdar@oracle.com" <subhra.mazumdar@oracle.com>,
        "dhaval.giani@oracle.com" <dhaval.giani@oracle.com>,
        "rohit.k.jain@oracle.com" <rohit.k.jain@oracle.com>,
        "daniel.m.jordan@oracle.com" <daniel.m.jordan@oracle.com>,
        "pavel.tatashin@microsoft.com" <pavel.tatashin@microsoft.com>,
        "matt@codeblueprint.co.uk" <matt@codeblueprint.co.uk>,
        "umgwanakikbuti@gmail.com" <umgwanakikbuti@gmail.com>,
        "riel@redhat.com" <riel@redhat.com>,
        "jbacik@fb.com" <jbacik@fb.com>,
        "juri.lelli@redhat.com" <juri.lelli@redhat.com>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        Ganapatrao Kulkarni <gkulkarni@marvell.com>,
        Jayachandran Chandrasekharan Nair <jnair@marvell.com>
Subject: Re: [PATCH 00/10] steal tasks to improve CPU utilization
Thread-Topic: [PATCH 00/10] steal tasks to improve CPU utilization
Thread-Index: AQHUgRpW+u9w4O+dU0K3NK6yx4LCBA==
Date:   Fri, 4 Jan 2019 13:37:17 +0000
Message-ID: <DM6PR18MB246096BE8C516A5A5D125932D98E0@DM6PR18MB2460.namprd18.prod.outlook.com>
References: <1540220381-424433-1-git-send-email-steven.sistare@oracle.com>
Accept-Language: en-US
Content-Language: en-US
received-spf: None (protection.outlook.com: marvell.com does not designate
 permitted sender hosts)
spamdiagnosticoutput: 1:99
spamdiagnosticmetadata: NSPM
Content-Type: text/plain; charset="us-ascii"
Content-Transfer-Encoding: quoted-printable
MIME-Version: 1.0
X-MS-Exchange-CrossTenant-Network-Message-Id: a2429633-a96f-4ea6-fc33-08d67249be9f
X-MS-Exchange-CrossTenant-originalarrivaltime: 04 Jan 2019 13:37:17.4432
 (UTC)
X-MS-Exchange-CrossTenant-fromentityheader: Hosted
X-MS-Exchange-CrossTenant-id: 70e1fb47-1155-421d-87fc-2e58f638b6e0
X-MS-Exchange-Transport-CrossTenantHeadersStamped: DM6PR18MB2716
X-OriginatorOrg: marvell.com
X-Proofpoint-Virus-Version: vendor=fsecure engine=2.50.10434:,, definitions=2019-01-04_05:,,
 signatures=0
X-Proofpoint-Details: rule=outbound_notspam policy=outbound score=0 priorityscore=1501
 malwarescore=0 suspectscore=0 phishscore=0 bulkscore=0 spamscore=0
 clxscore=1011 lowpriorityscore=0 mlxscore=0 impostorscore=0
 mlxlogscore=999 adultscore=0 classifier=spam adjust=0 reason=mlx
 scancount=1 engine=8.0.1-1810050000 definitions=main-1901040120
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk
List-ID: <linux-kernel.vger.kernel.org>
X-Mailing-List: linux-kernel@vger.kernel.org

On 22-Oct-18 8:40 PM, Steve Sistare wrote:=0A=
> =0A=
> When a CPU has no more CFS tasks to run, and idle_balance() fails to=0A=
> find a task, then attempt to steal a task from an overloaded CPU in the=
=0A=
> same LLC. Maintain and use a bitmap of overloaded CPUs to efficiently=0A=
> identify candidates.  To minimize search time, steal the first migratable=
=0A=
> task that is found when the bitmap is traversed.  For fairness, search=0A=
> for migratable tasks on an overloaded CPU in order of next to run.=0A=
> =0A=
> This simple stealing yields a higher CPU utilization than idle_balance()=
=0A=
> alone, because the search is cheap, so it may be called every time the CP=
U=0A=
> is about to go idle.  idle_balance() does more work because it searches=
=0A=
> widely for the busiest queue, so to limit its CPU consumption, it decline=
s=0A=
> to search if the system is too busy.  Simple stealing does not offload th=
e=0A=
> globally busiest queue, but it is much better than running nothing at all=
.=0A=
> =0A=
> The bitmap of overloaded CPUs is a new type of sparse bitmap, designed to=
=0A=
> reduce cache contention vs the usual bitmap when many threads concurrentl=
y=0A=
> set, clear, and visit elements.=0A=
> =0A=
> Patch 1 defines the sparsemask type and its operations.=0A=
> =0A=
> Patches 2, 3, and 4 implement the bitmap of overloaded CPUs.=0A=
> =0A=
> Patches 5 and 6 refactor existing code for a cleaner merge of later=0A=
>    patches.=0A=
> =0A=
> Patches 7 and 8 implement task stealing using the overloaded CPUs bitmap.=
=0A=
> =0A=
> Patch 9 disables stealing on systems with more than 2 NUMA nodes for the=
=0A=
> time being because of performance regressions that are not due to stealin=
g=0A=
> per-se.  See the patch description for details.=0A=
> =0A=
> Patch 10 adds schedstats for comparing the new behavior to the old, and=
=0A=
>    provided as a convenience for developers only, not for integration.=0A=
> =0A=
> The patch series is based on kernel 4.19.0-rc7.  It compiles, boots, and=
=0A=
> runs with/without each of CONFIG_SCHED_SMT, CONFIG_SMP, CONFIG_SCHED_DEBU=
G,=0A=
> and CONFIG_PREEMPT.  It runs without error with CONFIG_DEBUG_PREEMPT +=0A=
> CONFIG_SLUB_DEBUG + CONFIG_DEBUG_PAGEALLOC + CONFIG_DEBUG_MUTEXES +=0A=
> CONFIG_DEBUG_SPINLOCK + CONFIG_DEBUG_ATOMIC_SLEEP.  CPU hot plug and CPU=
=0A=
> bandwidth control were tested.=0A=
> =0A=
> Stealing imprroves utilization with only a modest CPU overhead in schedul=
er=0A=
> code.  In the following experiment, hackbench is run with varying numbers=
=0A=
> of groups (40 tasks per group), and the delta in /proc/schedstat is shown=
=0A=
> for each run, averaged per CPU, augmented with these non-standard stats:=
=0A=
> =0A=
>    %find - percent of time spent in old and new functions that search for=
=0A=
>      idle CPUs and tasks to steal and set the overloaded CPUs bitmap.=0A=
> =0A=
>    steal - number of times a task is stolen from another CPU.=0A=
> =0A=
> X6-2: 1 socket * 10 cores * 2 hyperthreads =3D 20 CPUs=0A=
> Intel(R) Xeon(R) CPU E5-2630 v4 @ 2.20GHz=0A=
> hackbench <grps> process 100000=0A=
> sched_wakeup_granularity_ns=3D15000000=0A=
> =0A=
>    baseline=0A=
>    grps  time  %busy  slice   sched   idle     wake %find  steal=0A=
>    1    8.084  75.02   0.10  105476  46291    59183  0.31      0=0A=
>    2   13.892  85.33   0.10  190225  70958   119264  0.45      0=0A=
>    3   19.668  89.04   0.10  263896  87047   176850  0.49      0=0A=
>    4   25.279  91.28   0.10  322171  94691   227474  0.51      0=0A=
>    8   47.832  94.86   0.09  630636 144141   486322  0.56      0=0A=
> =0A=
>    new=0A=
>    grps  time  %busy  slice   sched   idle     wake %find  steal  %speedu=
p=0A=
>    1    5.938  96.80   0.24   31255   7190    24061  0.63   7433  36.1=0A=
>    2   11.491  99.23   0.16   74097   4578    69512  0.84  19463  20.9=0A=
>    3   16.987  99.66   0.15  115824   1985   113826  0.77  24707  15.8=0A=
>    4   22.504  99.80   0.14  167188   2385   164786  0.75  29353  12.3=0A=
>    8   44.441  99.86   0.11  389153   1616   387401  0.67  38190   7.6=0A=
> =0A=
> Elapsed time improves by 8 to 36%, and CPU busy utilization is up=0A=
> by 5 to 22% hitting 99% for 2 or more groups (80 or more tasks).=0A=
> The cost is at most 0.4% more find time.=0A=
> =0A=
> Additional performance results follow.  A negative "speedup" is a=0A=
> regression.  Note: for all hackbench runs, sched_wakeup_granularity_ns=0A=
> is set to 15 msec.  Otherwise, preemptions increase at higher loads and=
=0A=
> distort the comparison between baseline and new.=0A=
> =0A=
> ------------------ 1 Socket Results ------------------=0A=
> =0A=
> X6-2: 1 socket * 10 cores * 2 hyperthreads =3D 20 CPUs=0A=
> Intel(R) Xeon(R) CPU E5-2630 v4 @ 2.20GHz=0A=
> Average of 10 runs of: hackbench <groups> process 100000=0A=
> =0A=
>              --- base --    --- new ---=0A=
>    groups    time %stdev    time %stdev  %speedup=0A=
>         1   8.008    0.1   5.905    0.2      35.6=0A=
>         2  13.814    0.2  11.438    0.1      20.7=0A=
>         3  19.488    0.2  16.919    0.1      15.1=0A=
>         4  25.059    0.1  22.409    0.1      11.8=0A=
>         8  47.478    0.1  44.221    0.1       7.3=0A=
> =0A=
> X6-2: 1 socket * 22 cores * 2 hyperthreads =3D 44 CPUs=0A=
> Intel(R) Xeon(R) CPU E5-2699 v4 @ 2.20GHz=0A=
> Average of 10 runs of: hackbench <groups> process 100000=0A=
> =0A=
>              --- base --    --- new ---=0A=
>    groups    time %stdev    time %stdev  %speedup=0A=
>         1   4.586    0.8   4.596    0.6      -0.3=0A=
>         2   7.693    0.2   5.775    1.3      33.2=0A=
>         3  10.442    0.3   8.288    0.3      25.9=0A=
>         4  13.087    0.2  11.057    0.1      18.3=0A=
>         8  24.145    0.2  22.076    0.3       9.3=0A=
>        16  43.779    0.1  41.741    0.2       4.8=0A=
> =0A=
> KVM 4-cpu=0A=
> Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz=0A=
> tbench, average of 11 runs.=0A=
> =0A=
>    clients    %speedup=0A=
>          1        16.2=0A=
>          2        11.7=0A=
>          4         9.9=0A=
>          8        12.8=0A=
>         16        13.7=0A=
> =0A=
> KVM 2-cpu=0A=
> Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz=0A=
> =0A=
>    Benchmark                     %speedup=0A=
>    specjbb2015_critical_jops          5.7=0A=
>    mysql_sysb1.0.14_mutex_2          40.6=0A=
>    mysql_sysb1.0.14_oltp_2            3.9=0A=
> =0A=
> ------------------ 2 Socket Results ------------------=0A=
> =0A=
> X6-2: 2 sockets * 10 cores * 2 hyperthreads =3D 40 CPUs=0A=
> Intel(R) Xeon(R) CPU E5-2630 v4 @ 2.20GHz=0A=
> Average of 10 runs of: hackbench <groups> process 100000=0A=
> =0A=
>              --- base --    --- new ---=0A=
>    groups    time %stdev    time %stdev  %speedup=0A=
>         1   7.945    0.2   7.219    8.7      10.0=0A=
>         2   8.444    0.4   6.689    1.5      26.2=0A=
>         3  12.100    1.1   9.962    2.0      21.4=0A=
>         4  15.001    0.4  13.109    1.1      14.4=0A=
>         8  27.960    0.2  26.127    0.3       7.0=0A=
> =0A=
> X6-2: 2 sockets * 22 cores * 2 hyperthreads =3D 88 CPUs=0A=
> Intel(R) Xeon(R) CPU E5-2699 v4 @ 2.20GHz=0A=
> Average of 10 runs of: hackbench <groups> process 100000=0A=
> =0A=
>              --- base --    --- new ---=0A=
>    groups    time %stdev    time %stdev  %speedup=0A=
>         1   5.826    5.4   5.840    5.0      -0.3=0A=
>         2   5.041    5.3   6.171   23.4     -18.4=0A=
>         3   6.839    2.1   6.324    3.8       8.1=0A=
>         4   8.177    0.6   7.318    3.6      11.7=0A=
>         8  14.429    0.7  13.966    1.3       3.3=0A=
>        16  26.401    0.3  25.149    1.5       4.9=0A=
> =0A=
> =0A=
> X6-2: 2 sockets * 22 cores * 2 hyperthreads =3D 88 CPUs=0A=
> Intel(R) Xeon(R) CPU E5-2699 v4 @ 2.20GHz=0A=
> Oracle database OLTP, logging disabled, NVRAM storage=0A=
> =0A=
>    Customers   Users   %speedup=0A=
>      1200000      40       -1.2=0A=
>      2400000      80        2.7=0A=
>      3600000     120        8.9=0A=
>      4800000     160        4.4=0A=
>      6000000     200        3.0=0A=
> =0A=
> X6-2: 2 sockets * 14 cores * 2 hyperthreads =3D 56 CPUs=0A=
> Intel(R) Xeon(R) CPU E5-2690 v4 @ 2.60GHz=0A=
> Results from the Oracle "Performance PIT".=0A=
> =0A=
>    Benchmark                                           %speedup=0A=
> =0A=
>    mysql_sysb1.0.14_fileio_56_rndrd                        19.6=0A=
>    mysql_sysb1.0.14_fileio_56_seqrd                        12.1=0A=
>    mysql_sysb1.0.14_fileio_56_rndwr                         0.4=0A=
>    mysql_sysb1.0.14_fileio_56_seqrewr                      -0.3=0A=
> =0A=
>    pgsql_sysb1.0.14_fileio_56_rndrd                        19.5=0A=
>    pgsql_sysb1.0.14_fileio_56_seqrd                         8.6=0A=
>    pgsql_sysb1.0.14_fileio_56_rndwr                         1.0=0A=
>    pgsql_sysb1.0.14_fileio_56_seqrewr                       0.5=0A=
> =0A=
>    opatch_time_ASM_12.2.0.1.0_HP2M                          7.5=0A=
>    select-1_users-warm_asmm_ASM_12.2.0.1.0_HP2M             5.1=0A=
>    select-1_users_asmm_ASM_12.2.0.1.0_HP2M                  4.4=0A=
>    swingbenchv3_asmm_soebench_ASM_12.2.0.1.0_HP2M           5.8=0A=
> =0A=
>    lm3_memlat_L2                                            4.8=0A=
>    lm3_memlat_L1                                            0.0=0A=
> =0A=
>    ub_gcc_56CPUs-56copies_Pipe-based_Context_Switching     60.1=0A=
>    ub_gcc_56CPUs-56copies_Shell_Scripts_1_concurrent        5.2=0A=
>    ub_gcc_56CPUs-56copies_Shell_Scripts_8_concurrent       -3.0=0A=
>    ub_gcc_56CPUs-56copies_File_Copy_1024_bufsize_2000_maxblocks 2.4=0A=
> =0A=
> X5-2: 2 sockets * 18 cores * 2 hyperthreads =3D 72 CPUs=0A=
> Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz=0A=
> =0A=
>    NAS_OMP=0A=
>    bench class   ncpu    %improved(Mops)=0A=
>    dc    B       72      1.3=0A=
>    is    C       72      0.9=0A=
>    is    D       72      0.7=0A=
> =0A=
>    sysbench mysql, average of 24 runs=0A=
>            --- base ---     --- new ---=0A=
>    nthr   events  %stdev   events  %stdev %speedup=0A=
>       1    331.0    0.25    331.0    0.24     -0.1=0A=
>       2    661.3    0.22    661.8    0.22      0.0=0A=
>       4   1297.0    0.88   1300.5    0.82      0.2=0A=
>       8   2420.8    0.04   2420.5    0.04     -0.1=0A=
>      16   4826.3    0.07   4825.4    0.05     -0.1=0A=
>      32   8815.3    0.27   8830.2    0.18      0.1=0A=
>      64  12823.0    0.24  12823.6    0.26      0.0=0A=
> =0A=
> --------------------------------------------------------------=0A=
> =0A=
> Steve Sistare (10):=0A=
>    sched: Provide sparsemask, a reduced contention bitmap=0A=
>    sched/topology: Provide hooks to allocate data shared per LLC=0A=
>    sched/topology: Provide cfs_overload_cpus bitmap=0A=
>    sched/fair: Dynamically update cfs_overload_cpus=0A=
>    sched/fair: Hoist idle_stamp up from idle_balance=0A=
>    sched/fair: Generalize the detach_task interface=0A=
>    sched/fair: Provide can_migrate_task_llc=0A=
>    sched/fair: Steal work from an overloaded CPU when CPU goes idle=0A=
>    sched/fair: disable stealing if too many NUMA nodes=0A=
>    sched/fair: Provide idle search schedstats=0A=
> =0A=
>   include/linux/sched/topology.h |   1 +=0A=
>   include/linux/sparsemask.h     | 260 +++++++++++++++++++++++++++++++=0A=
>   kernel/sched/core.c            |  30 +++-=0A=
>   kernel/sched/fair.c            | 338 ++++++++++++++++++++++++++++++++++=
+++----=0A=
>   kernel/sched/features.h        |   6 +=0A=
>   kernel/sched/sched.h           |  13 +-=0A=
>   kernel/sched/stats.c           |  11 +-=0A=
>   kernel/sched/stats.h           |  13 ++=0A=
>   kernel/sched/topology.c        | 117 +++++++++++++-=0A=
>   lib/Makefile                   |   2 +-=0A=
>   lib/sparsemask.c               | 142 +++++++++++++++++=0A=
>   11 files changed, 898 insertions(+), 35 deletions(-)=0A=
>   create mode 100644 include/linux/sparsemask.h=0A=
>   create mode 100644 lib/sparsemask.c=0A=
> =0A=
> --=0A=
> 1.8.3.1=0A=
> =0A=
> =0A=
=0A=
Hi Steve,=0A=
=0A=
Tried your patchset on ThunderX2 with 2 nodes. Please find my observations =
below.=0A=
=0A=
Hackbench was run on single node due to variance on 2 nodes and it showed =
=0A=
improvement under load.=0A=
=0A=
Single node hackbench numbers:=0A=
group    old time	new time	steals		%change=0A=
1	 6.717		7.275		21		-8.31=0A=
2	 8.449		9.268		106		-9.69=0A=
3	 12.035		12.761		173071		-6.03=0A=
4	 14.648		9.787		595889		 33.19=0A=
8	 22.513		18.329		2397394		 18.58=0A=
16	 39.861		36.263		3949903		 9.06=0A=
=0A=
column "new time" shows hackbench runtime in seconds with the patchset.=0A=
=0A=
Tried below benchmarks with 2 nodes, but no performance benefit/degradation=
 was =0A=
observed on multiple runs.=0A=
  - MySQL (read/write/PS etc with sysbench)=0A=
  - HHVM running oss-performance benchmarks=0A=
=0A=
Shijith=0A=