On Fri, May 31, 2019 at 4:16 AM Nitesh Narayan Lal <nitesh@xxxxxxxxxx> wrote: > > > On 5/30/19 5:53 PM, Alexander Duyck wrote: > > This series provides an asynchronous means of hinting to a hypervisor > > that a guest page is no longer in use and can have the data associated > > with it dropped. To do this I have implemented functionality that allows > > for what I am referring to as "waste page treatment". > > > > I have based many of the terms and functionality off of waste water > > treatment, the idea for the similarity occured to me after I had reached > > the point of referring to the hints as "bubbles", as the hints used the > > same approach as the balloon functionality but would disappear if they > > were touched, as a result I started to think of the virtio device as an > > aerator. The general idea with all of this is that the guest should be > > treating the unused pages so that when they end up heading "downstream" > > to either another guest, or back at the host they will not need to be > > written to swap. > > > > So for a bit of background for the treatment process, it is based on a > > sequencing batch reactor (SBR)[1]. The treatment process itself has five > > stages. The first stage is the fill, with this we take the raw pages and > > add them to the reactor. The second stage is react, in this stage we hand > > the pages off to the Virtio Balloon driver to have hints attached to them > > and for those hints to be sent to the hypervisor. The third stage is > > settle, in this stage we are waiting for the hypervisor to process the > > pages, and we should receive an interrupt when it is completed. The fourth > > stage is to decant, or drain the reactor of pages. Finally we have the > > idle stage which we will go into if the reference count for the reactor > > gets down to 0 after a drain, or if a fill operation fails to obtain any > > pages and the reference count has hit 0. Otherwise we return to the first > > state and start the cycle over again. > > > > This patch set is still far more intrusive then I would really like for > > what it has to do. Currently I am splitting the nr_free_pages into two > > values and having to add a pointer and an index to track where we area in > > the treatment process for a given free_area. I'm also not sure I have > > covered all possible corner cases where pages can get into the free_area > > or move from one migratetype to another. > > > > Also I am still leaving a number of things hard-coded such as limiting the > > lowest order processed to PAGEBLOCK_ORDER, and have left it up to the > > guest to determine what size of reactor it wants to allocate to process > > the hints. > > > > Another consideration I am still debating is if I really want to process > > the aerator_cycle() function in interrupt context or if I should have it > > running in a thread somewhere else. > > Can you please share some performance numbers? > > I will be sharing a less mm-intrusive bitmap-based approach hopefully by > next week. > Let's compare the two approaches then, in the meanwhile I will start > reviewing your patch-set. The performance can vary quite a bit depending on the configuration. So for example with the memory shuffling enabled I saw an overall improvement in transactions in the page_fault1 test I was running, however I suspect that is just due to the fact that I inlined the bit that was doing the shuffling at the 2nd patch in. I'm still working on gathering data so you can consider the data provided below as preliminary, and I want to emphasize your mileage may vary as it seems like the config used can make a big difference. So the results below are for a will-it-scale test of a VM running with 16 VCPUs and 32G of memory. The clean version is without patches applied, and "aerate" is with patches applied. I disabled the memory shuffling in the config for the kernels since it seemed like an unfair comparison with it enabled. Before the test I ran "memhog 32g" to pull in all available memory on the "clean" test and to pull in and flush all the memory on the "aerate" tests. One thing that isn't really making sense to me yet is why the results for the aerate version appear to be better then the clean version when we start getting into higher thread counts. One thing I notice is that clear_page_erms jumps to the top of a perf trace on the host at about the inflection point where the "clean" guest starts to under-perform versus the "aerate" guest. So it is possible that there may be some benefit to having the host clear the pages before the guest processes them. 5.2.0-rc2-next-20190529-clean #53 tasks,processes,processes_idle,threads,threads_idle,linear 0,0,100,0,100,0 1,574916,93.73,574313,93.70,574916 2,1006964,87.47,918228,87.52,1149832 3,1373857,81.23,1170468,82.35,1724748 4,1781250,74.98,1526831,76.77,2299664 5,1973790,68.74,1764815,69.86,2874580 6,2235127,62.53,1912371,65.42,3449496 7,2499192,56.28,1936901,61.63,4024412 8,2581220,50.05,2114032,56.54,4599328 9,2804630,43.81,2202166,52.37,5174244 10,2746340,37.58,2194305,48.31,5749160 11,2694687,31.33,2189028,41.74,6324076 12,2772102,25.16,2176312,40.85,6898992 13,2854235,18.94,2146288,37.61,7473908 14,2720456,12.73,2063334,32.67,8048824 15,2753005,6.51,2103228,26.65,8623740 16,2595824,0.36,2142308,25.96,9198656 tasks,processes,processes_idle,threads,threads_idle,linear 0,0,100,0,100,0 1,568948,93.73,570092,93.72,570092 2,1006781,87.47,911829,87.57,1140184 3,1360418,81.23,1189920,82.22,1710276 4,1749889,74.99,1476555,77.22,2280368 5,1927251,68.76,1681522,70.49,2850460 6,2221112,62.51,1845148,65.74,3420552 7,2497960,56.29,1983406,61.44,3990644 8,2586250,50.01,2062633,56.99,4560736 9,2570559,43.82,1989225,53.14,5130828 10,2692389,37.57,2159570,48.07,5700920 11,2621505,31.33,2214469,43.73,6271012 12,2772863,25.15,2164639,40.35,6841104 13,2839319,18.94,2184126,36.90,7411196 14,2712433,12.77,2048788,31.16,7981288 15,2779543,6.54,2105144,27.29,8551380 16,2605799,0.34,2101187,23.20,9121472 5.2.0-rc2-next-20190529-aerate+ #55 tasks,processes,processes_idle,threads,threads_idle,linear 0,0,100,0,100,0 1,538715,93.73,538909,93.73,538909 2,985096,87.46,899393,87.54,1077818 3,1421187,81.25,1271836,81.88,1616727 4,1745358,75.00,1435337,77.61,2155636 5,2031097,68.76,1766946,70.37,2694545 6,2234646,62.51,1794401,66.94,3233454 7,2455541,56.27,2101020,59.42,3772363 8,2576793,50.09,1810192,59.45,4311272 9,2772082,43.82,2315719,50.58,4850181 10,2794868,37.62,1996644,50.84,5389090 11,2931943,31.36,2147434,42.58,5927999 12,2837655,25.12,2032434,42.79,6466908 13,2881797,18.95,2163387,36.80,7005817 14,2802190,12.73,2049732,30.00,7544726 15,2684374,6.53,2039098,26.48,8083635 16,2695848,0.41,2044131,22.08,8622544 tasks,processes,processes_idle,threads,threads_idle,linear 0,0,100,0,100,0 1,533361,93.72,532973,93.73,533361 2,980085,87.46,904796,87.50,1066722 3,1387100,81.21,1271080,81.41,1600083 4,1720030,74.99,1539417,75.99,2133444 5,1942111,68.74,1530612,73.21,2666805 6,2226552,62.51,1777038,66.97,3200166 7,2469715,56.27,2084451,59.72,3733527 8,2567333,50.04,1820491,59.38,4266888 9,2744551,43.82,2259861,51.73,4800249 10,2768107,37.60,2240844,48.10,5333610 11,2879636,31.37,2134152,46.56,5866971 12,2826859,25.18,1960830,44.07,6400332 13,2905216,19.05,1887735,36.66,6933693 14,2841688,12.79,2047092,30.18,7467054 15,2832234,6.57,2066059,25.31,8000415 16,2758579,0.38,1961050,22.16,8533776
