On Mon, Jul 29, 2024 at 11:22 AM Huang, Ying <ying.huang@xxxxxxxxx> wrote: > > Hi, Yafang, > > Yafang Shao <laoar.shao@xxxxxxxxx> writes: > > > During my recent work to resolve latency spikes caused by zone->lock > > contention[0], I found that CONFIG_PCP_BATCH_SCALE_MAX is difficult to use > > in practice. > > As we discussed before [1], I still feel confusing about the description > about zone->lock contention. How about change the description to > something like, Sure, I will change it. > > Larger page allocation/freeing batch number may cause longer run time of > code holding zone->lock. If zone->lock is heavily contended at the same > time, latency spikes may occur even for casual page allocation/freeing. > Although reducing the batch number cannot make zone->lock contended > lighter, it can reduce the latency spikes effectively. > > [1] https://lore.kernel.org/linux-mm/87ttgv8hlz.fsf@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/ > > > To demonstrate this, I wrote a Python script: > > > > import mmap > > > > size = 6 * 1024**3 > > > > while True: > > mm = mmap.mmap(-1, size) > > mm[:] = b'\xff' * size > > mm.close() > > > > Run this script 10 times in parallel and measure the allocation latency by > > measuring the duration of rmqueue_bulk() with the BCC tools > > funclatency[1]: > > > > funclatency -T -i 600 rmqueue_bulk > > > > Here are the results for both AMD and Intel CPUs. > > > > AMD EPYC 7W83 64-Core Processor, single NUMA node, KVM virtual server > > ===================================================================== > > > > - Default value of 5 > > > > 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 : 12 | | > > 1024 -> 2047 : 9116 | | > > 2048 -> 4095 : 2004 | | > > 4096 -> 8191 : 2497 | | > > 8192 -> 16383 : 2127 | | > > 16384 -> 32767 : 2483 | | > > 32768 -> 65535 : 10102 | | > > 65536 -> 131071 : 212730 |******************* | > > 131072 -> 262143 : 314692 |***************************** | > > 262144 -> 524287 : 430058 |****************************************| > > 524288 -> 1048575 : 224032 |******************** | > > 1048576 -> 2097151 : 73567 |****** | > > 2097152 -> 4194303 : 17079 |* | > > 4194304 -> 8388607 : 3900 | | > > 8388608 -> 16777215 : 750 | | > > 16777216 -> 33554431 : 88 | | > > 33554432 -> 67108863 : 2 | | > > > > avg = 449775 nsecs, total: 587066511229 nsecs, count: 1305242 > > > > The avg alloc latency can be 449us, and the max latency can be higher > > than 30ms. > > > > - Value set to 0 > > > > 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 : 92 | | > > 1024 -> 2047 : 8594 | | > > 2048 -> 4095 : 2042818 |****** | > > 4096 -> 8191 : 8737624 |************************** | > > 8192 -> 16383 : 13147872 |****************************************| > > 16384 -> 32767 : 8799951 |************************** | > > 32768 -> 65535 : 2879715 |******** | > > 65536 -> 131071 : 659600 |** | > > 131072 -> 262143 : 204004 | | > > 262144 -> 524287 : 78246 | | > > 524288 -> 1048575 : 30800 | | > > 1048576 -> 2097151 : 12251 | | > > 2097152 -> 4194303 : 2950 | | > > 4194304 -> 8388607 : 78 | | > > > > avg = 19359 nsecs, total: 708638369918 nsecs, count: 36604636 > > > > The avg was reduced significantly to 19us, and the max latency is reduced > > to less than 8ms. > > > > - Conclusion > > > > On this AMD CPU, reducing vm.pcp_batch_scale_max significantly helps reduce > > latency. Latency-sensitive applications will benefit from this tuning. > > > > However, I don't have access to other types of AMD CPUs, so I was unable to > > test it on different AMD models. > > > > Intel(R) Xeon(R) Platinum 8260 CPU @ 2.40GHz, two NUMA nodes > > ============================================================ > > > > - Default value of 5 > > > > 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 : 2419 | | > > 1024 -> 2047 : 34499 |* | > > 2048 -> 4095 : 4272 | | > > 4096 -> 8191 : 9035 | | > > 8192 -> 16383 : 4374 | | > > 16384 -> 32767 : 2963 | | > > 32768 -> 65535 : 6407 | | > > 65536 -> 131071 : 884806 |****************************************| > > 131072 -> 262143 : 145931 |****** | > > 262144 -> 524287 : 13406 | | > > 524288 -> 1048575 : 1874 | | > > 1048576 -> 2097151 : 249 | | > > 2097152 -> 4194303 : 28 | | > > > > avg = 96173 nsecs, total: 106778157925 nsecs, count: 1110263 > > > > - Conclusion > > > > This Intel CPU works fine with the default setting. > > > > Intel(R) Xeon(R) Platinum 8260 CPU @ 2.40GHz, single NUMA node > > ============================================================== > > > > Using the cpuset cgroup, we can restrict the test script to run on NUMA > > node 0 only. > > > > - Default value of 5 > > > > 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 : 46 | | > > 512 -> 1023 : 695 | | > > 1024 -> 2047 : 19950 |* | > > 2048 -> 4095 : 1788 | | > > 4096 -> 8191 : 3392 | | > > 8192 -> 16383 : 2569 | | > > 16384 -> 32767 : 2619 | | > > 32768 -> 65535 : 3809 | | > > 65536 -> 131071 : 616182 |****************************************| > > 131072 -> 262143 : 295587 |******************* | > > 262144 -> 524287 : 75357 |**** | > > 524288 -> 1048575 : 15471 |* | > > 1048576 -> 2097151 : 2939 | | > > 2097152 -> 4194303 : 243 | | > > 4194304 -> 8388607 : 3 | | > > > > avg = 144410 nsecs, total: 150281196195 nsecs, count: 1040651 > > > > The zone->lock contention becomes severe when there is only a single NUMA > > node. The average latency is approximately 144us, with the maximum > > latency exceeding 4ms. > > > > - Value set to 0 > > > > 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 : 24 | | > > 512 -> 1023 : 2686 | | > > 1024 -> 2047 : 10246 | | > > 2048 -> 4095 : 4061529 |********* | > > 4096 -> 8191 : 16894971 |****************************************| > > 8192 -> 16383 : 6279310 |************** | > > 16384 -> 32767 : 1658240 |*** | > > 32768 -> 65535 : 445760 |* | > > 65536 -> 131071 : 110817 | | > > 131072 -> 262143 : 20279 | | > > 262144 -> 524287 : 4176 | | > > 524288 -> 1048575 : 436 | | > > 1048576 -> 2097151 : 8 | | > > 2097152 -> 4194303 : 2 | | > > > > avg = 8401 nsecs, total: 247739809022 nsecs, count: 29488508 > > > > After setting it to 0, the avg latency is reduced to around 8us, and the > > max latency is less than 4ms. > > > > - Conclusion > > > > On this Intel CPU, this tuning doesn't help much. Latency-sensitive > > applications work well with the default setting. > > > > It is worth noting that all the above data were tested using the upstream > > kernel. > > > > Why introduce a systl knob? > > =========================== > > > > From the above data, it's clear that different CPU types have varying > > allocation latencies concerning zone->lock contention. Typically, people > > don't release individual kernel packages for each type of x86_64 CPU. > > > > Furthermore, for latency-insensitive applications, we can keep the default > > setting for better throughput. In our production environment, we set this > > value to 0 for applications running on Kubernetes servers while keeping it > > at the default value of 5 for other applications like big data. It's not > > common to release individual kernel packages for each application. > > Thanks for detailed performance data! > > Is there any downside observed to set CONFIG_PCP_BATCH_SCALE_MAX to 0 in > your environment? If not, I suggest to use 0 as default for > CONFIG_PCP_BATCH_SCALE_MAX. Because we have clear evidence that > CONFIG_PCP_BATCH_SCALE_MAX hurts latency for some workloads. After > that, if someone found some other workloads need larger > CONFIG_PCP_BATCH_SCALE_MAX, we can make it tunable dynamically. > The decision doesn’t rest with us, the kernel team at our company. It’s made by the system administrators who manage a large number of servers. The latency spikes only occur on the Kubernetes (k8s) servers, not in other environments like big data servers. We have informed other system administrators, such as those managing the big data servers, about the latency spike issues, but they are unwilling to make the change. No one wants to make changes unless there is evidence showing that the old settings will negatively impact them. However, as you know, latency is not a critical concern for big data; throughput is more important. If we keep the current settings, we will have to release different kernel packages for different environments, which is a significant burden for us. -- Regards Yafang
