2020년 4월 9일 (목) 오전 1:55, Vlastimil Babka <vbabka@xxxxxxx>님이 작성: > > On 4/3/20 7:40 AM, js1304@xxxxxxxxx wrote: > > From: Joonsoo Kim <iamjoonsoo.kim@xxxxxxx> > > > > Hello, > > > > This patchset implements workingset protection and detection on > > the anonymous LRU list. > > Hi! Hi! > > I did another test to show the performance effect of this patchset. > > > > - ebizzy (with modified random function) > > ebizzy is the test program that main thread allocates lots of memory and > > child threads access them randomly during the given times. Swap-in/out > > will happen if allocated memory is larger than the system memory. > > > > The random function that represents the zipf distribution is used to > > make hot/cold memory. Hot/cold ratio is controlled by the parameter. If > > the parameter is high, hot memory is accessed much larger than cold one. > > If the parameter is low, the number of access on each memory would be > > similar. I uses various parameters in order to show the effect of > > patchset on various hot/cold ratio workload. > > > > My test setup is a virtual machine with 8 cpus and 1024MB memory. > > > > Result format is as following. > > > > Parameter 0.1 ... 1.3 > > Allocated memory size > > Throughput for base (larger is better) > > Throughput for patchset (larger is better) > > Improvement (larger is better) > > > > > > * single thread > > > > 0.1 0.3 0.5 0.7 0.9 1.1 1.3 > > <512M> > > 7009.0 7372.0 7774.0 8523.0 9569.0 10724.0 11936.0 > > 6973.0 7342.0 7745.0 8576.0 9441.0 10730.0 12033.0 > > -0.01 -0.0 -0.0 0.01 -0.01 0.0 0.01 > > <768M> > > 915.0 1039.0 1275.0 1687.0 2328.0 3486.0 5445.0 > > 920.0 1037.0 1238.0 1689.0 2384.0 3638.0 5381.0 > > 0.01 -0.0 -0.03 0.0 0.02 0.04 -0.01 > > <1024M> > > 425.0 471.0 539.0 753.0 1183.0 2130.0 3839.0 > > 414.0 468.0 553.0 770.0 1242.0 2187.0 3932.0 > > -0.03 -0.01 0.03 0.02 0.05 0.03 0.02 > > <1280M> > > 320.0 346.0 410.0 556.0 871.0 1654.0 3298.0 > > 316.0 346.0 411.0 550.0 892.0 1652.0 3293.0 > > -0.01 0.0 0.0 -0.01 0.02 -0.0 -0.0 > > <1536M> > > 273.0 290.0 341.0 458.0 733.0 1381.0 2925.0 > > 271.0 293.0 344.0 462.0 740.0 1398.0 2969.0 > > -0.01 0.01 0.01 0.01 0.01 0.01 0.02 > > <2048M> > > 77.0 79.0 95.0 147.0 276.0 690.0 1816.0 > > 91.0 94.0 115.0 170.0 321.0 770.0 2018.0 > > 0.18 0.19 0.21 0.16 0.16 0.12 0.11 > > > > > > * multi thread (8) > > > > 0.1 0.3 0.5 0.7 0.9 1.1 1.3 > > <512M> > > 29083.0 29648.0 30145.0 31668.0 33964.0 38414.0 43707.0 > > 29238.0 29701.0 30301.0 31328.0 33809.0 37991.0 43667.0 > > 0.01 0.0 0.01 -0.01 -0.0 -0.01 -0.0 > > <768M> > > 3332.0 3699.0 4673.0 5830.0 8307.0 12969.0 17665.0 > > 3579.0 3992.0 4432.0 6111.0 8699.0 12604.0 18061.0 > > 0.07 0.08 -0.05 0.05 0.05 -0.03 0.02 > > <1024M> > > 1921.0 2141.0 2484.0 3296.0 5391.0 8227.0 14574.0 > > 1989.0 2155.0 2609.0 3565.0 5463.0 8170.0 15642.0 > > 0.04 0.01 0.05 0.08 0.01 -0.01 0.07 > > <1280M> > > 1524.0 1625.0 1931.0 2581.0 4155.0 6959.0 12443.0 > > 1560.0 1707.0 2016.0 2714.0 4262.0 7518.0 13910.0 > > 0.02 0.05 0.04 0.05 0.03 0.08 0.12 > > <1536M> > > 1303.0 1399.0 1550.0 2137.0 3469.0 6712.0 12944.0 > > 1356.0 1465.0 1701.0 2237.0 3583.0 6830.0 13580.0 > > 0.04 0.05 0.1 0.05 0.03 0.02 0.05 > > <2048M> > > 172.0 184.0 215.0 289.0 514.0 1318.0 4153.0 > > 175.0 190.0 225.0 329.0 606.0 1585.0 5170.0 > > 0.02 0.03 0.05 0.14 0.18 0.2 0.24 > > > > As we can see, as allocated memory grows, patched kernel get the better > > result. Maximum improvement is 21% for the single thread test and > > 24% for the multi thread test. > > So, these results seem to be identical since v1. After the various changes up to > v5, should the benchmark be redone? And was that with a full patchset or > patches 1+2? It was done with a full patchset. I think that these results would not be changed even on v5 since it is improvement from the concept of this patchset and implementation detail doesn't much matter. However, I will redo. > > * EXPERIMENT > > I made a test program to imitates above scenario and confirmed that > > problem exists. Then, I checked that this patchset fixes it. > > > > My test setup is a virtual machine with 8 cpus and 6100MB memory. But, > > the amount of the memory that the test program can use is about 280 MB. > > This is because the system uses large ram-backed swap and large ramdisk > > to capture the trace. > > > > Test scenario is like as below. > > > > 1. allocate cold memory (512MB) > > 2. allocate hot-1 memory (96MB) > > 3. activate hot-1 memory (96MB) > > 4. allocate another hot-2 memory (96MB) > > 5. access cold memory (128MB) > > 6. access hot-2 memory (96MB) > > 7. repeat 5, 6 > > > > Since hot-1 memory (96MB) is on the active list, the inactive list can > > contains roughly 190MB pages. hot-2 memory's re-access interval > > (96+128 MB) is more 190MB, so it cannot be promoted without workingset > > detection and swap-in/out happens repeatedly. With this patchset, > > workingset detection works and promotion happens. Therefore, swap-in/out > > occurs less. > > > > Here is the result. (average of 5 runs) > > > > type swap-in swap-out > > base 863240 989945 > > patch 681565 809273 > > > > As we can see, patched kernel do less swap-in/out. > > Same comment, also v1 has this note: I had tested this scenario on every version of the patchset and found the same trend. > > Note that, this result is gotten from v5.1. Although workingset detection > > works on v5.1, it doesn't work well on v5.5. It looks like that recent > > code change on workingset.c is the reason of this problem. I will > > track it soon. > What was the problem then, assuming it's fixed? Maybe I just missed it > mentioned. Can results now be gathered on 5.6? It was fixed on v2. Change log on v2 "fix a critical bug that uses out of index lru list in workingset_refault()" is for this problem. I should note that clearly. > In general, this patchset seems to be doing the right thing. I haven't reviewed > the code yet, but hope to do so soon. But inevitably, with any changes in this > area there will be workloads that will suffer instead of benefit. That can be > because we are actually doing a wrong thing, or there's a bug in the code, or > the workloads happen to benefit from the current behavior even if it's not the > generally optimal one. And I'm afraid only testing on a variety of workloads can > show that. You mentioned somewhere that your production workloads benefit? Can > it be quantified more? Could e.g. Johannes test this a bit at Facebook, or I cannot share the detail of the test for my production (smart TV) workload. Roughly, it is repeat of various action and app (channel change, volume change, youtube, etc.) on smart TV and it is memory stress test. Result after the workload is: base pswpin 328211 pswpout 304015 patched pswpin 261884 pswpout 276062 So, improvement on pswpin and pswpout is roughly 20% and 9%, respectively. > it be quantified more? Could e.g. Johannes test this a bit at Facebook, or > somebody at Google? It's really helpful if someone else could test this on their workload. Thanks.
