Yafang Shao <laoar.shao@xxxxxxxxx> writes: > On Thu, Jul 11, 2024 at 7:05 PM Huang, Ying <ying.huang@xxxxxxxxx> wrote: >> >> Yafang Shao <laoar.shao@xxxxxxxxx> writes: >> >> > On Thu, Jul 11, 2024 at 4:38 PM Huang, Ying <ying.huang@xxxxxxxxx> wrote: >> >> >> >> Yafang Shao <laoar.shao@xxxxxxxxx> writes: >> >> >> >> > On Thu, Jul 11, 2024 at 2:40 PM Huang, Ying <ying.huang@xxxxxxxxx> wrote: >> >> >> >> >> >> Yafang Shao <laoar.shao@xxxxxxxxx> writes: >> >> >> >> >> >> > On Wed, Jul 10, 2024 at 11:02 AM Huang, Ying <ying.huang@xxxxxxxxx> wrote: >> >> >> >> >> >> >> >> Yafang Shao <laoar.shao@xxxxxxxxx> writes: >> >> >> >> >> >> >> >> > Background >> >> >> >> > ========== >> >> >> >> > >> >> >> >> > In our containerized environment, we have a specific type of container >> >> >> >> > that runs 18 processes, each consuming approximately 6GB of RSS. These >> >> >> >> > processes are organized as separate processes rather than threads due >> >> >> >> > to the Python Global Interpreter Lock (GIL) being a bottleneck in a >> >> >> >> > multi-threaded setup. Upon the exit of these containers, other >> >> >> >> > containers hosted on the same machine experience significant latency >> >> >> >> > spikes. >> >> >> >> > >> >> >> >> > Investigation >> >> >> >> > ============= >> >> >> >> > >> >> >> >> > My investigation using perf tracing revealed that the root cause of >> >> >> >> > these spikes is the simultaneous execution of exit_mmap() by each of >> >> >> >> > the exiting processes. This concurrent access to the zone->lock >> >> >> >> > results in contention, which becomes a hotspot and negatively impacts >> >> >> >> > performance. The perf results clearly indicate this contention as a >> >> >> >> > primary contributor to the observed latency issues. >> >> >> >> > >> >> >> >> > + 77.02% 0.00% uwsgi [kernel.kallsyms] [k] mmput >> >> >> >> > - 76.98% 0.01% uwsgi [kernel.kallsyms] [k] exit_mmap >> >> >> >> > - 76.97% exit_mmap >> >> >> >> > - 58.58% unmap_vmas >> >> >> >> > - 58.55% unmap_single_vma >> >> >> >> > - unmap_page_range >> >> >> >> > - 58.32% zap_pte_range >> >> >> >> > - 42.88% tlb_flush_mmu >> >> >> >> > - 42.76% free_pages_and_swap_cache >> >> >> >> > - 41.22% release_pages >> >> >> >> > - 33.29% free_unref_page_list >> >> >> >> > - 32.37% free_unref_page_commit >> >> >> >> > - 31.64% free_pcppages_bulk >> >> >> >> > + 28.65% _raw_spin_lock >> >> >> >> > 1.28% __list_del_entry_valid >> >> >> >> > + 3.25% folio_lruvec_lock_irqsave >> >> >> >> > + 0.75% __mem_cgroup_uncharge_list >> >> >> >> > 0.60% __mod_lruvec_state >> >> >> >> > 1.07% free_swap_cache >> >> >> >> > + 11.69% page_remove_rmap >> >> >> >> > 0.64% __mod_lruvec_page_state >> >> >> >> > - 17.34% remove_vma >> >> >> >> > - 17.25% vm_area_free >> >> >> >> > - 17.23% kmem_cache_free >> >> >> >> > - 17.15% __slab_free >> >> >> >> > - 14.56% discard_slab >> >> >> >> > free_slab >> >> >> >> > __free_slab >> >> >> >> > __free_pages >> >> >> >> > - free_unref_page >> >> >> >> > - 13.50% free_unref_page_commit >> >> >> >> > - free_pcppages_bulk >> >> >> >> > + 13.44% _raw_spin_lock >> >> >> >> >> >> >> >> I don't think your change will reduce zone->lock contention cycles. So, >> >> >> >> I don't find the value of the above data. >> >> >> >> >> >> >> >> > By enabling the mm_page_pcpu_drain() we can locate the pertinent page, >> >> >> >> > with the majority of them being regular order-0 user pages. >> >> >> >> > >> >> >> >> > <...>-1540432 [224] d..3. 618048.023883: mm_page_pcpu_drain: page=0000000035a1b0b7 pfn=0x11c19c72 order=0 migratetyp >> >> >> >> > e=1 >> >> >> >> > <...>-1540432 [224] d..3. 618048.023887: <stack trace> >> >> >> >> > => free_pcppages_bulk >> >> >> >> > => free_unref_page_commit >> >> >> >> > => free_unref_page_list >> >> >> >> > => release_pages >> >> >> >> > => free_pages_and_swap_cache >> >> >> >> > => tlb_flush_mmu >> >> >> >> > => zap_pte_range >> >> >> >> > => unmap_page_range >> >> >> >> > => unmap_single_vma >> >> >> >> > => unmap_vmas >> >> >> >> > => exit_mmap >> >> >> >> > => mmput >> >> >> >> > => do_exit >> >> >> >> > => do_group_exit >> >> >> >> > => get_signal >> >> >> >> > => arch_do_signal_or_restart >> >> >> >> > => exit_to_user_mode_prepare >> >> >> >> > => syscall_exit_to_user_mode >> >> >> >> > => do_syscall_64 >> >> >> >> > => entry_SYSCALL_64_after_hwframe >> >> >> >> > >> >> >> >> > The servers experiencing these issues are equipped with impressive >> >> >> >> > hardware specifications, including 256 CPUs and 1TB of memory, all >> >> >> >> > within a single NUMA node. The zoneinfo is as follows, >> >> >> >> > >> >> >> >> > Node 0, zone Normal >> >> >> >> > pages free 144465775 >> >> >> >> > boost 0 >> >> >> >> > min 1309270 >> >> >> >> > low 1636587 >> >> >> >> > high 1963904 >> >> >> >> > spanned 564133888 >> >> >> >> > present 296747008 >> >> >> >> > managed 291974346 >> >> >> >> > cma 0 >> >> >> >> > protection: (0, 0, 0, 0) >> >> >> >> > ... >> >> >> >> > pagesets >> >> >> >> > cpu: 0 >> >> >> >> > count: 2217 >> >> >> >> > high: 6392 >> >> >> >> > batch: 63 >> >> >> >> > vm stats threshold: 125 >> >> >> >> > cpu: 1 >> >> >> >> > count: 4510 >> >> >> >> > high: 6392 >> >> >> >> > batch: 63 >> >> >> >> > vm stats threshold: 125 >> >> >> >> > cpu: 2 >> >> >> >> > count: 3059 >> >> >> >> > high: 6392 >> >> >> >> > batch: 63 >> >> >> >> > >> >> >> >> > ... >> >> >> >> > >> >> >> >> > The pcp high is around 100 times the batch size. >> >> >> >> > >> >> >> >> > I also traced the latency associated with the free_pcppages_bulk() >> >> >> >> > function during the container exit process: >> >> >> >> > >> >> >> >> > nsecs : count distribution >> >> >> >> > 0 -> 1 : 0 | | >> >> >> >> > 2 -> 3 : 0 | | >> >> >> >> > 4 -> 7 : 0 | | >> >> >> >> > 8 -> 15 : 0 | | >> >> >> >> > 16 -> 31 : 0 | | >> >> >> >> > 32 -> 63 : 0 | | >> >> >> >> > 64 -> 127 : 0 | | >> >> >> >> > 128 -> 255 : 0 | | >> >> >> >> > 256 -> 511 : 148 |***************** | >> >> >> >> > 512 -> 1023 : 334 |****************************************| >> >> >> >> > 1024 -> 2047 : 33 |*** | >> >> >> >> > 2048 -> 4095 : 5 | | >> >> >> >> > 4096 -> 8191 : 7 | | >> >> >> >> > 8192 -> 16383 : 12 |* | >> >> >> >> > 16384 -> 32767 : 30 |*** | >> >> >> >> > 32768 -> 65535 : 21 |** | >> >> >> >> > 65536 -> 131071 : 15 |* | >> >> >> >> > 131072 -> 262143 : 27 |*** | >> >> >> >> > 262144 -> 524287 : 84 |********** | >> >> >> >> > 524288 -> 1048575 : 203 |************************ | >> >> >> >> > 1048576 -> 2097151 : 284 |********************************** | >> >> >> >> > 2097152 -> 4194303 : 327 |*************************************** | >> >> >> >> > 4194304 -> 8388607 : 215 |************************* | >> >> >> >> > 8388608 -> 16777215 : 116 |************* | >> >> >> >> > 16777216 -> 33554431 : 47 |***** | >> >> >> >> > 33554432 -> 67108863 : 8 | | >> >> >> >> > 67108864 -> 134217727 : 3 | | >> >> >> >> > >> >> >> >> > The latency can reach tens of milliseconds. >> >> >> >> > >> >> >> >> > Experimenting >> >> >> >> > ============= >> >> >> >> > >> >> >> >> > vm.percpu_pagelist_high_fraction >> >> >> >> > -------------------------------- >> >> >> >> > >> >> >> >> > The kernel version currently deployed in our production environment is the >> >> >> >> > stable 6.1.y, and my initial strategy involves optimizing the >> >> >> >> >> >> >> >> IMHO, we should focus on upstream activity in the cover letter and patch >> >> >> >> description. And I don't think that it's necessary to describe the >> >> >> >> alternative solution with too much details. >> >> >> >> >> >> >> >> > vm.percpu_pagelist_high_fraction parameter. By increasing the value of >> >> >> >> > vm.percpu_pagelist_high_fraction, I aim to diminish the batch size during >> >> >> >> > page draining, which subsequently leads to a substantial reduction in >> >> >> >> > latency. After setting the sysctl value to 0x7fffffff, I observed a notable >> >> >> >> > improvement in latency. >> >> >> >> > >> >> >> >> > nsecs : count distribution >> >> >> >> > 0 -> 1 : 0 | | >> >> >> >> > 2 -> 3 : 0 | | >> >> >> >> > 4 -> 7 : 0 | | >> >> >> >> > 8 -> 15 : 0 | | >> >> >> >> > 16 -> 31 : 0 | | >> >> >> >> > 32 -> 63 : 0 | | >> >> >> >> > 64 -> 127 : 0 | | >> >> >> >> > 128 -> 255 : 120 | | >> >> >> >> > 256 -> 511 : 365 |* | >> >> >> >> > 512 -> 1023 : 201 | | >> >> >> >> > 1024 -> 2047 : 103 | | >> >> >> >> > 2048 -> 4095 : 84 | | >> >> >> >> > 4096 -> 8191 : 87 | | >> >> >> >> > 8192 -> 16383 : 4777 |************** | >> >> >> >> > 16384 -> 32767 : 10572 |******************************* | >> >> >> >> > 32768 -> 65535 : 13544 |****************************************| >> >> >> >> > 65536 -> 131071 : 12723 |************************************* | >> >> >> >> > 131072 -> 262143 : 8604 |************************* | >> >> >> >> > 262144 -> 524287 : 3659 |********** | >> >> >> >> > 524288 -> 1048575 : 921 |** | >> >> >> >> > 1048576 -> 2097151 : 122 | | >> >> >> >> > 2097152 -> 4194303 : 5 | | >> >> >> >> > >> >> >> >> > However, augmenting vm.percpu_pagelist_high_fraction can also decrease the >> >> >> >> > pcp high watermark size to a minimum of four times the batch size. While >> >> >> >> > this could theoretically affect throughput, as highlighted by Ying[0], we >> >> >> >> > have yet to observe any significant difference in throughput within our >> >> >> >> > production environment after implementing this change. >> >> >> >> > >> >> >> >> > Backporting the series "mm: PCP high auto-tuning" >> >> >> >> > ------------------------------------------------- >> >> >> >> >> >> >> >> Again, not upstream activity. We can describe the upstream behavior >> >> >> >> directly. >> >> >> > >> >> >> > Andrew has requested that I provide a more comprehensive analysis of >> >> >> > this issue, and in response, I have endeavored to outline all the >> >> >> > pertinent details in a thorough and detailed manner. >> >> >> >> >> >> IMHO, upstream activity can provide comprehensive analysis of the issue >> >> >> too. And, your patch has changed much from the first version. It's >> >> >> better to describe your current version. >> >> > >> >> > After backporting the pcp auto-tuning feature to the 6.1.y branch, the >> >> > code is almost the same with the upstream kernel wrt the pcp. I have >> >> > thoroughly documented the detailed data showcasing the changes in the >> >> > backported version, providing a clear picture of the results. However, >> >> > it's crucial to note that I am unable to directly run the upstream >> >> > kernel on our production environment due to practical constraints. >> >> >> >> IMHO, the patch is for upstream kernel, not some downstream kernel, so >> >> focus should be the upstream activity. The issue of the upstream >> >> kernel, and how to resolve it. The production environment test results >> >> can be used to support the upstream change. >> > >> > The sole distinction in the pcp between version 6.1.y and the >> > upstream kernel lies solely in the modifications made to the code by >> > you. Furthermore, given that your code changes have now been >> > successfully backported, what else do you expect me to do ? >> >> If you can run the upstream kernel directly with some proxy workloads, >> it will be better. But, I understand that this may be not easy for you. >> >> So, what I really expect you to do is to organize the patch description >> in an upstream centric way. Describe the issue of the upstream kernel, >> and how do you resolve it. Although your test data comes from a >> downstream kernel with the same page allocator behavior. >> >> >> >> >> >> >> > My second endeavor was to backport the series titled >> >> >> >> > "mm: PCP high auto-tuning"[1], which comprises nine individual patches, >> >> >> >> > into our 6.1.y stable kernel version. Subsequent to its deployment in our >> >> >> >> > production environment, I noted a pronounced reduction in latency. The >> >> >> >> > observed outcomes are as enumerated below: >> >> >> >> > >> >> >> >> > nsecs : count distribution >> >> >> >> > 0 -> 1 : 0 | | >> >> >> >> > 2 -> 3 : 0 | | >> >> >> >> > 4 -> 7 : 0 | | >> >> >> >> > 8 -> 15 : 0 | | >> >> >> >> > 16 -> 31 : 0 | | >> >> >> >> > 32 -> 63 : 0 | | >> >> >> >> > 64 -> 127 : 0 | | >> >> >> >> > 128 -> 255 : 0 | | >> >> >> >> > 256 -> 511 : 0 | | >> >> >> >> > 512 -> 1023 : 0 | | >> >> >> >> > 1024 -> 2047 : 2 | | >> >> >> >> > 2048 -> 4095 : 11 | | >> >> >> >> > 4096 -> 8191 : 3 | | >> >> >> >> > 8192 -> 16383 : 1 | | >> >> >> >> > 16384 -> 32767 : 2 | | >> >> >> >> > 32768 -> 65535 : 7 | | >> >> >> >> > 65536 -> 131071 : 198 |********* | >> >> >> >> > 131072 -> 262143 : 530 |************************ | >> >> >> >> > 262144 -> 524287 : 824 |************************************** | >> >> >> >> > 524288 -> 1048575 : 852 |****************************************| >> >> >> >> > 1048576 -> 2097151 : 714 |********************************* | >> >> >> >> > 2097152 -> 4194303 : 389 |****************** | >> >> >> >> > 4194304 -> 8388607 : 143 |****** | >> >> >> >> > 8388608 -> 16777215 : 29 |* | >> >> >> >> > 16777216 -> 33554431 : 1 | | >> >> >> >> > >> >> >> >> > Compared to the previous data, the maximum latency has been reduced to >> >> >> >> > less than 30ms. >> >> >> >> >> >> >> >> People don't care too much about page freeing latency during processes >> >> >> >> exiting. Instead, they care more about the process exiting time, that >> >> >> >> is, throughput. So, it's better to show the page allocation latency >> >> >> >> which is affected by the simultaneous processes exiting. >> >> >> > >> >> >> > I'm confused also. Is this issue really hard to understand ? >> >> >> >> >> >> IMHO, it's better to prove the issue directly. If you cannot prove it >> >> >> directly, you can try alternative one and describe why. >> >> > >> >> > Not all data can be verified straightforwardly or effortlessly. The >> >> > primary focus lies in the zone->lock contention, which necessitates >> >> > measuring the latency it incurs. To accomplish this, the >> >> > free_pcppages_bulk() function serves as an effective tool for >> >> > evaluation. Therefore, I have opted to specifically measure the >> >> > latency associated with free_pcppages_bulk(). >> >> > >> >> > The rationale behind not measuring allocation latency is due to the >> >> > necessity of finding a willing participant to endure potential delays, >> >> > a task that proved unsuccessful as no one expressed interest. In >> >> > contrast, assessing free_pcppages_bulk()'s latency solely requires >> >> > identifying and experimenting with the source causing the delays, >> >> > making it a more feasible approach. >> >> >> >> Can you run a benchmark program that do quite some memory allocation by >> >> yourself to test it? >> > >> > I can have a try. >> >> Thanks! >> >> > However, is it the key point here? >> >> It's better to prove the issue directly instead of indirectly. >> >> > Why can't the lock contention be measured by the freeing? >> >> Have you measured the lock contention after adjusting >> CONFIG_PCP_BATCH_SCALE_MAX? IIUC, the lock contention will become even >> worse. Smaller CONFIG_PCP_BATCH_SCALE_MAX helps latency, but it will >> hurt lock contention. I have said it several times, but it seems that >> you don't agree with me. Can you prove I'm wrong with data? > > Now I understand the point. It seems we have different understandings > regarding the zone lock contention. > > CPU A (Freer) CPU B (Allocator) > lock zone->lock > free pages lock zone->lock > unlock zone->lock alloc pages > unlock zone->lock > > If the Freer holds the zone lock for an extended period, the Allocator > has to wait, right? Isn't that a lock contention issue? Lock > contention affects not only CPU system usage but also latency. Thanks for explanation! Now, I understand our difference. When I measure the spin lock contention, I usually measure lock spinning CPU cycles. The more spinning cycles, the more severe the lock contention. IIUC, we cannot measure the lock contention with the latency only. For example, if there are only 1 "Freer" and 0 "Allocator" in the system, there will be no lock contention at all. But, if we increase page free batch number, the latency of "Freer" and "Allocator" will increase without the lock contention. Another example, the total spin cycles decrease, but the spin cycles of each waiting increase because the total number of waiting decreases with larger batch number. So, the lock contention becomes better at the cost of latency. So, I suggest to use the latency spike to describe the issue and the solution. I suggest you to measure the processes exit time too. To evaluate the possible negative impact on throughput of smaller batch number. -- Best Regards, Huang, Ying
