Very sorry, I forgot to send my last response as plain text. > On Apr 3, 2018, at 6:31 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote: > > On Mon 02-04-18 09:24:22, Buddy Lumpkin wrote: >> Page replacement is handled in the Linux Kernel in one of two ways: >> >> 1) Asynchronously via kswapd >> 2) Synchronously, via direct reclaim >> >> At page allocation time the allocating task is immediately given a page >> from the zone free list allowing it to go right back to work doing >> whatever it was doing; Probably directly or indirectly executing business >> logic. >> >> Just prior to satisfying the allocation, free pages is checked to see if >> it has reached the zone low watermark and if so, kswapd is awakened. >> Kswapd will start scanning pages looking for inactive pages to evict to >> make room for new page allocations. The work of kswapd allows tasks to >> continue allocating memory from their respective zone free list without >> incurring any delay. >> >> When the demand for free pages exceeds the rate that kswapd tasks can >> supply them, page allocation works differently. Once the allocating task >> finds that the number of free pages is at or below the zone min watermark, >> the task will no longer pull pages from the free list. Instead, the task >> will run the same CPU-bound routines as kswapd to satisfy its own >> allocation by scanning and evicting pages. This is called a direct reclaim. >> >> The time spent performing a direct reclaim can be substantial, often >> taking tens to hundreds of milliseconds for small order0 allocations to >> half a second or more for order9 huge-page allocations. In fact, kswapd is >> not actually required on a linux system. It exists for the sole purpose of >> optimizing performance by preventing direct reclaims. >> >> When memory shortfall is sufficient to trigger direct reclaims, they can >> occur in any task that is running on the system. A single aggressive >> memory allocating task can set the stage for collateral damage to occur in >> small tasks that rarely allocate additional memory. Consider the impact of >> injecting an additional 100ms of latency when nscd allocates memory to >> facilitate caching of a DNS query. >> >> The presence of direct reclaims 10 years ago was a fairly reliable >> indicator that too much was being asked of a Linux system. Kswapd was >> likely wasting time scanning pages that were ineligible for eviction. >> Adding RAM or reducing the working set size would usually make the problem >> go away. Since then hardware has evolved to bring a new struggle for >> kswapd. Storage speeds have increased by orders of magnitude while CPU >> clock speeds stayed the same or even slowed down in exchange for more >> cores per package. This presents a throughput problem for a single >> threaded kswapd that will get worse with each generation of new hardware. > > AFAIR we used to scale the number of kswapd workers many years ago. It > just turned out to be not all that great. We have a kswapd reclaim > window for quite some time and that can allow to tune how much proactive > kswapd should be. Are you referring to vm.watermark_scale_factor? This helps quite a bit. Previously I had to increase min_free_kbytes in order to get a larger gap between the low and min watemarks. I was very excited when saw that this had been added upstream. > > Also please note that the direct reclaim is a way to throttle overly > aggressive memory consumers. I totally agree, in fact I think this should be the primary role of direct reclaims because they have a substantial impact on performance. Direct reclaims are the emergency brakes for page allocation, and the case I am making here is that they used to only occur when kswapd had to skip over a lot of pages. This changed over time as the rate a system can allocate pages increased. Direct reclaims slowly became a normal part of page replacement. > The more we do in the background context > the easier for them it will be to allocate faster. So I am not really > sure that more background threads will solve the underlying problem. It > is just a matter of memory hogs tunning to end in the very same > situtation AFAICS. Moreover the more they are going to allocate the more > less CPU time will _other_ (non-allocating) task get. The important thing to realize here is that kswapd and direct reclaims run the same code paths. There is very little that they do differently. If you compare my test results with one kswapd vs four, your an see that direct reclaims increase the kernel mode CPU consumption considerably. By dedicating more threads to proactive page replacement, you eliminate direct reclaims which reduces the total number of parallel threads that are spinning on the CPU. > >> Test Details > > I will have to study this more to comment. > > [...] >> By increasing the number of kswapd threads, throughput increased by ~50% >> while kernel mode CPU utilization decreased or stayed the same, likely due >> to a decrease in the number of parallel tasks at any given time doing page >> replacement. > > Well, isn't that just an effect of more work being done on behalf of > other workload that might run along with your tests (and which doesn't > really need to allocate a lot of memory)? In other words how > does the patch behaves with a non-artificial mixed workloads? It works quite well. We are just starting to test our production apps. I will have results to share soon. > > Please note that I am not saying that we absolutely have to stick with the > current single-thread-per-node implementation but I would really like to > see more background on why we should be allowing heavy memory hogs to > allocate faster or how to prevent that. My test results demonstrate the problem very well. It shows that a handful of SSDs can create enough demand for kswapd that it consumes ~100% CPU long before throughput is able to reach it’s peak. Direct reclaims start occurring at that point. Aggregate throughput continues to increase, but eventually the pauses generated by the direct reclaims cause throughput to plateau: Test #3: 1 kswapd thread per node dd sy dd_cpu kswapd0 kswapd1 throughput dr pgscan_kswapd pgscan_direct 10 4 26.07 28.56 27.03 7355924.40 0 459316976 0 16 7 34.94 69.33 69.66 10867895.20 0 872661643 0 22 10 36.03 93.99 99.33 13130613.60 489 1037654473 11268334 28 10 30.34 95.90 98.60 14601509.60 671 1182591373 15429142 34 14 34.77 97.50 99.23 16468012.00 10850 1069005644 249839515 40 17 36.32 91.49 97.11 17335987.60 18903 975417728 434467710 46 19 38.40 90.54 91.61 17705394.40 25369 855737040 582427973 52 22 40.88 83.97 83.70 17607680.40 31250 709532935 724282458 58 25 40.89 82.19 80.14 17976905.60 35060 657796473 804117540 64 28 41.77 73.49 75.20 18001910.00 39073 561813658 895289337 70 33 45.51 63.78 64.39 17061897.20 44523 379465571 1020726436 76 36 46.95 57.96 60.32 16964459.60 47717 291299464 1093172384 82 39 47.16 55.43 56.16 16949956.00 49479 247071062 1134163008 88 42 47.41 53.75 47.62 16930911.20 51521 195449924 1180442208 90 43 47.18 51.40 50.59 16864428.00 51618 190758156 1183203901 I think we have reached the point where it makes sense for page replacement to have more than one mode. Enterprise class servers with lots of memory and a large number of CPU cores would benefit heavily if more threads could be devoted toward proactive page replacement. The polar opposite case is my Raspberry PI which I want to run as efficiently as possible. This problem is only going to get worse. I think it makes sense to be able to choose between efficiency and performance (throughput and latency reduction). > I would be also very interested > to see how to scale the number of threads based on how CPUs are utilized > by other workloads. > -- > Michal Hocko > SUSE Labs I agree. I think it would be nice to have a work queue that can sense when CPU utilization for a task peaks at 100% and uses that as criteria to start another task up to some maximum that was determined at boot time. I would also determine a max gap size for the watermarks at boot time as well, specifically the gap between min and low since that provides the buffer that absorbs spikey reclaim behavior as free pages drops. Each time an direct reclaim occurs, increase the gap up to the limit. Make the limit tunable as well. If at any time along the way CPU peaks at 100%, start another thread up to the limit established at boot (which is also tunable).
