Re: [RFC PATCH V1 0/6] sched/numa: Enhance disjoint VMA scanning

[Raghavendra K T <raghavendra.kt@amd.com>]

On 9/13/2023 10:58 AM, Swapnil Sapkal wrote:
> Hello Raghu,
> 
> On 8/29/2023 11:36 AM, Raghavendra K T wrote:
>> Since commit fc137c0ddab2 ("sched/numa: enhance vma scanning logic") [1]
>> VMA scanning is allowed if:
>> 1) The task had accessed the VMA.
>>   Rationale: Reduce overhead for the tasks that had not
>> touched VMA. Also filter out unnecessary scanning.
>>
>> 2) Early phase of the VMA scan where mm->numa_scan_seq is less than 2.
>>   Rationale: Understanding initial characteristics of VMAs and also
>>   prevent VMA scanning unfairness.
>>
>> While that works for most of the times to reduce scanning overhead,
>>   there are some corner cases associated with it.
>>
>> This was found in an internal LKP run and also reported by [2]. There was
>> an attempt to fix.
>>
>> Link: 
>> https://lore.kernel.org/linux-mm/cover.1685506205.git.raghavendra.kt@amd.com/T/ 
>>
>>
>> This is a fully different series after Mel's feedback to address the 
>> issue
>>   and also a continuation of enhancing VMA scanning for NUMA balancing.
>>
>> Problem statement (Disjoint VMA set):
>> ======================================
>> Let's look at some of the corner cases with a below example of tasks 
>> and their
>> access pattern.
>>
>> Consider N tasks (threads) of a process.
>> Set1 tasks accessing vma_x (group of VMAs)
>> Set2 tasks accessing vma_y (group of VMAs)
>>
>>               Set1                      Set2
>>          -------------------         --------------------
>>          | task_1..task_n/2 |       | task_n/2+1..task_n |
>>          -------------------         --------------------
>>                   |                             |
>>                   V                             V
>>          -------------------         --------------------
>>          |     vma_x       |         |     vma_y         |
>>          -------------------         --------------------
>>
>> Corner cases:
>> (a) Out of N tasks, not all of them gets fair opportunity to scan. 
>> (PeterZ).
>> suppose Set1 tasks gets more opportunity to scan (May be because of the
>> activity pattern of tasks or other reasons in current design) in the 
>> above
>> example, then vma_x gets scanned more number of times than vma_y.
>>
>> some experiment is also done here which illustrates this unfairness:
>> Link: 
>> https://lore.kernel.org/lkml/c730dee0-a711-8a8e-3eb1-1bfdd21e6add@amd.com/ 
>>
>>
>> (b) Sizes of vmas can differ.
>> Suppose size of vma_y is far greater than the size of vma_x, then a 
>> bigger
>> portion of vma_y can potentially be left unscanned since scanning is 
>> bounded
>> by scan_size of 256MB (default) for each iteration.
>>
>> (c) Highly active threads trap a few VMAs frequently, and some of the 
>> VMAs not
>> accessed for long time can potentially get starved of scanning 
>> indefinitely
>> (Mel). There is a possibility of lack of enough hints/details about 
>> VMAs if it
>> is needed later for migration.
>>
>> (d) Allocation of memory in some specific manner (Mel).
>> One example could be, Suppose a main thread allocates memory and it is 
>> not
>> active. When other threads tries to act upon it, they may not have much
>> hints about it, if the corresponding VMA was not scanned.
>>
>> (e) VMAs that are created after two full scans of mm 
>> (mm->numa_scan_seq > 2)
>> will never get scanned. (Observed rarely but very much possible 
>> depending on
>> workload behaviour).
>>
>> Above this, a combination of some of the above (e.g., (a) and (b)) can
>> potentially amplifyi/worsen the side effect.
>>
>> This patchset, tries to address the above issues by enhancing 
>> unconditional
>> VMA scanning logic.
>>
>> High level ideas:
>> =================
>> Idea-1) Depending on vma_size, populate a per vma_scan_select value, 
>> decrement it
>> and when it hits zero do force scan (Mel).
>> vma_scan_select value is again repopulated when it hits zero.
>>
>> This is how VMA scanning phases looks like after implementation:
>>
>> |<---p1--->|<-----p2----->|<-----p2----->|...
>>
>> Algorithm:
>> p1: New VMA, initial phase do not scan till scan_delay.
>>
>> p2: Allow scanning if the task has accessed VMA or vma_scan_select hit 
>> zero.
>>
>> Reinitialize vma_scan_select and repeat p2.
>>
>> pros/cons:
>> +  : Ratelimiting is inbuilt to the approach
>> +  : vma_size is taken into account for scanning
>> +/-: Scanning continues forever
>> -  : Changes in vma size is taken care after force scan. i.e.,
>>     vma_scan_select is repopulated only after vma_scan_select hits zero.
>>
>> Idea-1 can potentially cover all the issues mentioned above.
>>
>> Idea-2) Take bitmask_weight of latest access_pids value (suggested by 
>> Bharata).
>> If number of tasks accessing vma is >= 1, unconditionally allow scanning.
>>
>> Idea-3 ) Take bitmask_weight of access_pid history of VMA. If number 
>> of tasks
>> accessing VMA is > THRESHOLD (=3), unconditionally allow scanning.
>>
>> Rationale (Idea-2,3): Do not miss out scanning of critical VMAs.
>>
>> Idea-4) Have a per vma_scan_seq. allow the unconditional scan till 
>> vma_scan_seq
>> reaches a value proportional (or equal) to vma_size/scan_size.
>> This a complimentary to Idea-1.
>>
>> this is how VMA scanning phases looks like after implementation:
>>
>> |<--p1--->|<-----p2----->|<-----p3----->|<-----p4----->...||<-----p2----->|<-----p3----->|<-----p4-----> 
>> ...||
>>                                                          
>> RESET                                               RESET
>> Algorithm:
>> p1: New VMA, initial phase do not scan till scan_delay.
>>
>> p2: Allow scanning if task has accessed VMA or vma_scan_seq has 
>> reached till
>>   f(vma_size)/scan_size) for e.g., f = 1/2 * vma_size/scan_size.
>>
>> p3: Allow scanning if task has accessed VMA or vma_scan_seq has 
>> reached till
>>   f(vma_size)/scan_size in a rate limited manner. This is an optional 
>> phase.
>>
>> p4: Allow scanning iff task has accessed VMA.
>>
>> Reset after p4 (optional).
>>
>> Repeat p2, p3 p4
>>
>> Motivation: Allow agressive scanning in the beginning followed by a rate
>> limited scanning. And then completely disallow scanning to avoid 
>> unnecessary
>> scanning. Reset time could be a function of scan_delay and chosen long 
>> enough
>> to aid long running task to forget history and start afresh.
>>
>> +  : Ratelimiting need to be taken care separately if needed.
>> +/-: Scanning continues only if RESET of vma_scan_seq is implemented.
>> +  : changes in vma size is taken care in every scan.
>>
>>   Current patch series implements Ideas 1, 2, 3 + extension of access 
>> PID history
>> idea from PeterZ.
>>
>> Results:
>> ======
>> Base: 6.5.0-rc6+ (4853c74bd7ab)
>> SUT: Milan w/ 2 numa nodes 256 cpus
>>
>> mmtest        numa01_THREAD_ALLOC manual run:
>>
>>         base        patched
>> real        1m22.758s    1m9.200s
>> user        249m49.540s    229m30.039s
>> sys        0m25.040s    3m10.451s
>>
>> numa_pte_updates     6985    1573363
>> numa_hint_faults     2705    1022623
>> numa_hint_faults_local     2279    389633
>> numa_pages_migrated     426    632990
>>
>> kernbench
>>             base            patched
>> Amean     user-256    21989.09 (   0.00%)    21677.36 *   1.42%*
>> Amean     syst-256    10171.34 (   0.00%)    10818.28 *  -6.36%*
>> Amean     elsp-256      166.81 (   0.00%)      168.40 *  -0.95%*
>>
>> Duration User       65973.18    65038.00
>> Duration System     30538.92    32478.59
>> Duration Elapsed      529.52      533.09
>>
>> Ops NUMA PTE updates                  976844.00      962680.00
>> Ops NUMA hint faults                  226763.00      245620.00
>> Ops NUMA pages migrated               220146.00      207025.00
>> Ops AutoNUMA cost                       1144.84        1238.77
>>
>> Improvements in other benchmarks I have tested.
>> Time based:
>> Hashjoin    4.21%
>> Btree         2.04%
>> XSbench        0.36%
>>
>> Throughput based:
>> Graph500     -3.62%
>> Nas.bt        3.69%
>> Nas.ft        21.91%
>>
>> Note: VMA scanning improvements [1] has refined scanning so much that
>> system overhead we re-introduce with additional scan look glaringly
>> high. But If we consider the difference between before [1] and current
>> series, overall scanning overhead is considerably reduced.
>>
>> 1. Link: 
>> https://lore.kernel.org/lkml/cover.1677672277.git.raghavendra.kt@amd.com/T/#t 
>>
>> 2. Link: 
>> https://lore.kernel.org/lkml/cover.1683033105.git.raghavendra.kt@amd.com/
>>
>> Note: Patch description is again repeated in some patches to avoid any
>> need to copy from cover letter again.
>>
>> Peter Zijlstra (1):
>>    sched/numa: Increase tasks' access history
>>
>> Raghavendra K T (5):
>>    sched/numa: Move up the access pid reset logic
>>    sched/numa: Add disjoint vma unconditional scan logic
>>    sched/numa: Remove unconditional scan logic using mm numa_scan_seq
>>    sched/numa: Allow recently accessed VMAs to be scanned
>>    sched/numa: Allow scanning of shared VMAs
>>
>>   include/linux/mm.h       |  12 +++--
>>   include/linux/mm_types.h |   5 +-
>>   kernel/sched/fair.c      | 109 ++++++++++++++++++++++++++++++++-------
>>   3 files changed, 102 insertions(+), 24 deletions(-)
>>
> 
> I have tested this series on 4th generation EPYC processor. I am seeing 
> improvement in autonuma-benchmark with the series.
> 
> o System Details
> 
> - 4th Generation EPYC System
> - 2 x 128C/256T
> - NPS1 mode
> 
> o Kernels
> 
> base:   4853c74bd7ab Merge tag 'parisc-for-6.5-rc7' of 
> git://git.kernel.org/pub/scm/linux/kernel/git/deller/parisc-linux
> 
> 
> ==================================================================
> Test          : autonuma-benchmark
> Units         : Time in seconds
> Interpretation: Lower is better
> Statistic     : AMean
> ==================================================================
> commit:
>    base (4853c74bd7ab)
>    base + this_series
> 
>    base (4853c74bd7ab)         base + this_series
> ---------------- ---------------------------
>           %stddev     %change         %stddev
>               \          |                \
>      522.58           -11.2%     464.23        
> autonuma-benchmark.numa01.seconds
>      273.93            -1.2%     270.75        
> autonuma-benchmark.numa01_THREAD_ALLOC.seconds
>        0.60            +0.0%       0.60        
> autonuma-benchmark.numa02.seconds
>      102.68            +3.7%     106.50        
> autonuma-benchmark.numa02_SMT.seconds
> 
> Tested-by: Swapnil Sapkal <Swapnil.Sapkal@amd.com>
> 


Thank you Swapnil.

Regards
- Raghu