lkml.org 
[lkml]   [2019]   [Jul]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [RFC][PATCH v11 0/2] mm: Support for page hinting
From
Date

On 7/11/19 10:58 AM, Alexander Duyck wrote:
> On Thu, Jul 11, 2019 at 4:31 AM Nitesh Narayan Lal <nitesh@redhat.com> wrote:
>>
>> On 7/10/19 7:40 PM, Alexander Duyck wrote:
>>> On Wed, Jul 10, 2019 at 12:52 PM Nitesh Narayan Lal <nitesh@redhat.com> wrote:
>>>
>>> The results up here were redundant with what is below so I am just
>>> dropping them. I would suggest only including one set of results in
>>> any future cover page as it is confusing to duplicate it like that.
>>>
>>>> This approach tracks all freed pages of the order MAX_ORDER - 2 in bitmaps.
>>>> A new hook after buddy merging is used to set the bits in the bitmap.
>>>> Currently, the bits are only cleared when pages are hinted, not when pages are
>>>> re-allocated.
>>>>
>>>> Bitmaps are stored on a per-zone basis and are protected by the zone lock. A
>>>> workqueue asynchronously processes the bitmaps as soon as a pre-defined memory
>>>> threshold is met, trying to isolate and report pages that are still free.
>>>>
>>>> The isolated pages are reported via virtio-balloon, which is responsible for
>>>> sending batched pages to the host synchronously. Once the hypervisor processed
>>>> the hinting request, the isolated pages are returned back to the buddy.
>>>>
>>>> Changelog in v11:
>>>> * Added logic to take care of multiple NUMA nodes scenarios.
>>>> * Simplified the logic for reporting isolated pages to the host. (Eg. replaced
>>>> dynamically allocated arrays with static ones, introduced wait event instead of
>>>> the loop in order to wait for a response from the host)
>>>> * Added a mutex to prevent race condition when page hinting is enabled by
>>>> multiple drivers.
>>>> * Simplified the logic responsible for decrementing free page counter for each
>>>> zone.
>>>> * Simplified code structuring/naming.
>>>>
>>>> Known work items for the future:
>>>> * Test device assigned guests to ensure that hinting doesn't break it.
>>>> * Follow up on VIRTIO_BALLOON_F_PAGE_POISON's device-side support.
>>>> * Decide between MADV_DONTNEED and MADV_FREE.
>>>> * Look into memory hotplug, more efficient locking, better naming conventions to
>>>> avoid confusion with VIRTIO_BALLOON_F_FREE_PAGE_HINT support.
>>>> * Come up with proper/traceable error-message/logs and look into other code
>>>> simplifications. (If necessary).
>>>>
>>>> Benefit analysis:
>>>> 1. Number of 5GB guests (each touching 4GB memory) that can be launched without
>>>> swap usage on a system with 15GB:
>>>> unmodified kernel - 2, 3rd with 2.5GB
>>>> v11 page hinting - 6, 7th with 26MB
>>>> v1 bubble hinting - 6, 7th with 1.8GB
>>>>
>>>> Conclusion - In this particular testcase on using v11 page hinting and
>>>> v1 bubble-hinting 4 more guests could be launched without swapping compared
>>>> to an unmodified kernel.
>>>> For the 7th guest launch, v11 page hinting is slightly better than v1 Bubble
>>>> hinting as it touches lesser swap space.
>>> I'm confused by the comment. From what I can tell bubble hinting came
>>> up with 1.8GB of memory while page hinting only managed to achieve
>>> .026GB (Using the same units makes it easier to visualize the
>>> difference). Also your test says "can be launched without swap usage",
>>> yet you say the bubble hinting is touching swap which makes not sense
>>> to me.
>> I will work on the cover to improve this part.
>> Basically, In each case, the first number indicates the number of the
>> guest which are launched without touching the swap space. For instance
>> with bubble hinting, I was able to launch 6 guests without any swap
>> usage. On launching the 7th guests initially there was no swap usage,
>> however, as the test app starts allocating 4GB memory the swap came into
>> the picture. 1.8 GB is the swap usage after the completion of the test
>> application.
>>>> Setup & procedure -
>>>> Total NUMA Node Memory ~ 15 GB (All guests are run on a single NUMA node)
>>>> Guest Memory = 5GB
>>>> Number of CPUs in the guest = 1
>>>> Host swap = 4GB
>>>> Workload = test allocation program that allocates 4GB memory, touches it via
>>>> memset and exits.
>>>> The first guest is launched and once its console is up, the test allocation
>>>> program is executed with 4 GB memory request (Due to this the guest occupies
>>>> almost 4-5 GB of memory in the host in a system without page hinting). Once
>>>> this program exits at that time another guest is launched in the host and the
>>>> same process is followed. It is continued until the swap is not used.
>>>>
>>>> 2. Memhog execution time (For 3 guests each of 6GB on a system with 15GB):
>>>> unmodified kernel - Guest1:21s, Guest2:27s, Guest3:2m37s swap used = 3.7GB
>>>> v11 page hinting - Guest1:23s, Guest2:26s, Guest3:21s swap used = 0
>>>> v1 bubble hinting - Guest1:23, Guest2:11s, Guest3:26s swap used = 0
>>>>
>>>> For this particular test-case in a guest which doesn't require swap access
>>>> "memhog 6G" execution time lies within a range of 15-30s.
>>>> Conclusion -
>>>> In the above test case for an unmodified kernel on executing memhog in the
>>>> third guest execution time rises to above 2minutes due to swap access.
>>>> Using either page-hinting or bubble hinting brings this execution time to a
>>>> a normal range of 15-30s.
>>> So really this test doesn't add much in value. The whole reason why
>>> Guest3 runs so much slower is because it is going to swap. I initially
>>> did this to demonstrate a point, but now running this test doesn't
>>> prove much as it isn't really meant to be a performance test. It is
>>> essentially just a duplicate of the "how many guests can you run" test
>>> that is passing itself off as some sort of performance test.
>>>
>>> We could probably just drop this from future version of this as long
>>> as we verify that the memory hinting is freeing most of the memory
>>> back and the guest is reporting a size less than the total guest
>>> memory size.
>>>
>> +1, makes sense to keep just one of the above two.
>>>> Setup & procedure -
>>>> Total NUMA Node Memory ~ 15 GB (All guests are run on a single NUMA node)
>>>> Guest Memory = 6GB
>>>> Number of CPUs in the guest = 4
>>>> Process = 3 Guests are launched and the ‘memhog 6G’ execution time is monitored
>>>> one after the other in each of them.
>>>> Host swap = 4GB
>>>>
>>>> Performance Analysis:
>>>> 1. will-it-scale's page_faul1
>>>> Setup -
>>>> Guest Memory = 6GB
>>>> Number of cores = 24
>>>>
>>>> Unmodified kernel -
>>>> 0,0,100,0,100,0
>>>> 1,514453,95.84,519502,95.83,519502
>>>> 2,991485,91.67,932268,91.68,1039004
>>>> 3,1381237,87.36,1264214,87.64,1558506
>>>> 4,1789116,83.36,1597767,83.88,2078008
>>>> 5,2181552,79.20,1889489,80.08,2597510
>>>> 6,2452416,75.05,2001879,77.10,3117012
>>>> 7,2671047,70.90,2263866,73.22,3636514
>>>> 8,2930081,66.75,2333813,70.60,4156016
>>>> 9,3126431,62.60,2370108,68.28,4675518
>>>> 10,3211937,58.44,2454093,65.74,5195020
>>>> 11,3162172,54.32,2450822,63.21,5714522
>>>> 12,3154261,50.14,2272290,58.98,6234024
>>>> 13,3115174,46.02,2369679,57.74,6753526
>>>> 14,3150511,41.86,2470837,54.02,7273028
>>>> 15,3134158,37.71,2428129,51.98,7792530
>>>> 16,3143067,33.57,2340469,49.54,8312032
>>>> 17,3112457,29.43,2263627,44.81,8831534
>>>> 18,3089724,25.29,2181879,38.69,9351036
>>>> 19,3076878,21.15,2236505,40.01,9870538
>>>> 20,3091978,16.95,2266327,35.00,10390040
>>>> 21,3082927,12.84,2172578,28.12,10909542
>>>> 22,3055282,8.73,2176269,29.14,11429044
>>>> 23,3081144,4.56,2138442,24.87,11948546
>>>> 24,3075509,0.45,2173753,21.62,12468048
>>>>
>>>> page hinting -
>>>> 0,0,100,0,100,0
>>>> 1,491683,95.83,494366,95.82,494366
>>>> 2,988415,91.67,919660,91.68,988732
>>>> 3,1344829,87.52,1244608,87.69,1483098
>>>> 4,1797933,83.37,1625797,83.70,1977464
>>>> 5,2179009,79.21,1881534,80.13,2471830
>>>> 6,2449858,75.07,2078137,76.82,2966196
>>>> 7,2732122,70.90,2178105,73.75,3460562
>>>> 8,2910965,66.75,2340901,70.28,3954928
>>>> 9,3006665,62.61,2353748,67.91,4449294
>>>> 10,3164752,58.46,2377936,65.08,4943660
>>>> 11,3234846,54.32,2510149,63.14,5438026
>>>> 12,3165477,50.17,2412007,59.91,5932392
>>>> 13,3141457,46.05,2421548,57.85,6426758
>>>> 14,3135839,41.90,2378021,53.81,6921124
>>>> 15,3109113,37.75,2269290,51.76,7415490
>>>> 16,3093613,33.62,2346185,48.73,7909856
>>>> 17,3086542,29.49,2352140,46.19,8404222
>>>> 18,3048991,25.36,2217144,41.52,8898588
>>>> 19,2965500,21.18,2313614,38.18,9392954
>>>> 20,2928977,17.05,2175316,35.67,9887320
>>>> 21,2896667,12.91,2141311,28.90,10381686
>>>> 22,3047782,8.76,2177664,28.24,10876052
>>>> 23,2994503,4.58,2160976,22.97,11370418
>>>> 24,3038762,0.47,2053533,22.39,11864784
>>>>
>>>> bubble-hinting v1 -
>>>> 0,0,100,0,100,0
>>>> 1,515272,95.83,492355,95.81,515272
>>>> 2,985903,91.66,919653,91.68,1030544
>>>> 3,1475300,87.51,1353723,87.65,1545816
>>>> 4,1783938,83.36,1586307,83.78,2061088
>>>> 5,2093307,79.20,1867395,79.95,2576360
>>>> 6,2441370,75.05,2055421,76.65,3091632
>>>> 7,2650471,70.89,2246014,72.93,3606904
>>>> 8,2926782,66.75,2333601,70.41,4122176
>>>> 9,3107617,62.60,2383112,68.46,4637448
>>>> 10,3192332,58.44,2441626,65.84,5152720
>>>> 11,3268043,54.32,2235964,62.92,5667992
>>>> 12,3191105,50.18,2449045,60.49,6183264
>>>> 13,3145317,46.05,2377317,57.80,6698536
>>>> 14,3161552,41.91,2395814,53.26,7213808
>>>> 15,3140443,37.77,2333200,51.42,7729080
>>>> 16,3130866,33.65,2150967,46.11,8244352
>>>> 17,3112894,29.52,2372068,45.93,8759624
>>>> 18,3078424,25.39,2336211,39.85,9274896
>>>> 19,3036457,21.27,2224821,35.25,9790168
>>>> 20,3046330,17.13,2199755,37.43,10305440
>>>> 21,2981130,12.98,2214862,28.67,10820712
>>>> 22,3017481,8.84,2195996,29.69,11335984
>>>> 23,2979906,4.68,2173395,25.90,11851256
>>>> 24,2971170,0.52,2134311,21.89,12366528
>>> Okay, so this doesn't match up with the results you gave me last time
>>> (https://lore.kernel.org/lkml/afac6f92-74f5-4580-0303-12b7374e5011@redhat.com/),
>>> and actually more closely matches what I was expecting to see. The
>>> bubble-hinting patches are performing within a few percent of what the
>>> baseline kernel was doing.
>> Interestingly even with an unmodified kernel with every fresh boot, I
>> observed a certain amount of variability in the results which I stated
>> below.
>>> I am assuming the results from before had
>>> some additional debugging enabled for the bubble-hinting test that
>>> wasn't enabled for the other ones.
>> Nope, I had debugging options enabled for all the cases. This time
>> around I disabled all the debug options.
> We can agree to disagree I guess. Those debugging options had reduced
> the throughput by over 30% on the guest kernel in my test runs. I was
> never able to reproduce the data you reported as enabling the same
> debug features on an unmodified kernel had reduced the throughput for
> the test just the same as it did for the bubble hinting version. Were
> you running the debug options on the host kernel or the guest?
In the guest. Do the results which I shared without debug options, match
with what you have?
I am also curious to know if you see any variability in the results of
page_fault1 for an unmodified kernel with every fresh boot? If so how
often?
> I
> suppose it is possible that having those debug options enabled on the
> host might trigger similar behavior to what you reported since you
> were using MADV_FREE versus MADV_DONTNEED so you wouldn't have to
> reallocate the pages and could circumvent the page allocation
> debugging.
>
>>>> Conclusion -
>>>> For an unmodified kernel, with every fresh boot, there is 3-4% delta observed
>>>> in the results wrt the numbers mentioned above. For both bubble-hinting and
>>>> page-hinting, there was no noticeable degradation observed other than the
>>>> expected variability mentioned earlier.
>>>>
>>>> Page hinting vs bubble hinting:
>>>> From the benefits and performance perspective, both solutions look quite similar
>>>> so far. However, unlike bubble-hinting which is more invasive, the overall core
>>>> mm changes required for page hinting are minimal.
>>>>
>>>> [1] https://lkml.org/lkml/2019/6/19/926
>>> So I think I called it out in the review of the patch but I think we
>>> may want to see what happens if we increase the size of the memory in
>>> the guest to something more like 64G or larger. My main concern is
>>> that as we increase the size of memory the walk through the bitmap is
>>> going to become more and more expensive and I am worried that at some
>>> point it will start impacting the results.
>> Ok, I can try that scenario.
>>
>>
>>
>> --
>> Thanks
>> Nitesh
>>
--
Thanks
Nitesh

\
 
 \ /
  Last update: 2019-07-11 17:04    [W:0.082 / U:9.680 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site