lkml.org 
[lkml]   [2014]   [Oct]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRE: [PATCH RFC 0/5] sched,numa: task placement with complex NUMA topologies
Date
>This patch set integrates two algorithms I have previously tested,
>one for glueless mesh NUMA topologies, where NUMA nodes communicate
>with far-away nodes through intermediary nodes, and backplane
>topologies, where communication with far-away NUMA nodes happens
>through backplane controllers (which cannot run tasks).
>
>Due to the inavailability of 8 node systems, and the fact that I
>am flying out to Linuxcon Europe / Plumbers / KVM Forum on Friday,
>I have not tested these patches yet. However, with a conference (and
>many familiar faces) coming up, it seemed like a good idea to get
>the code out there, anyway.
>
>The algorithms have been tested before, on both kinds of system.
>The new thing about this series is that both algorithms have been
>integrated into the same code base, and new code to select the
>preferred_nid for tasks in numa groups.
>
>Placement of tasks on smaller, directly connected, NUMA systems
>should not be affected at all by this patch series.
>
>I am interested in reviews, as well as test results on larger
>NUMA systems :)


Tested-by: Chegu Vinod <chegu_vinod@hp.com>

---

Hi Rik,

Applied your RFC patches along with the patch : https://lkml.org/lkml/2014/10/9/604
on a 3.17.0 kernel on ran a quick test on two platforms with different NUMA topologies:

1) 8 socket Westmere-EX system (HT-off)
2) 8 socket IvyBridge-EX prototype system (HT-on)

On both these system ran 4 instances of a 2-socket wide SpecJbb2005 workload and
then ran 2 instances of a 4-socket wide SpecJbb2005 workload. Repeated the
experiment 10 times.

Having these patches enabled the desired NUMA nodes to selected for both sized
workloads on both these systems...which is quite an encouraging sign !

Pl. see below a sampling of memory consumed by the java processes (as reported
by numastat captured 120 seconds into each of the runs).

Thanks
Vinod

---


1) 8 socket Westmere-EX system with the following SLIT table :


node distances:
node 0 1 2 3 4 5 6 7
0: 10 12 17 17 19 19 19 19
1: 12 10 17 17 19 19 19 19
2: 17 17 10 12 19 19 19 19
3: 17 17 12 10 19 19 19 19
4: 19 19 19 19 10 12 17 17
5: 19 19 19 19 12 10 17 17
6: 19 19 19 19 17 17 10 12
7: 19 19 19 19 17 17 12 10

1a) 4 x 2 socket wide instances (each instance has 20 warehouse threads)

Note: The correct NUMA nodes got picked 9 out of 10 runs..

PID Node 0 Node 1 Node 2 Node 3 Node 4 Node 5 Node 6 Node 7 Total
15758 (java) 0 33 1904 1542 0 9 0 14 3503
15759 (java) 1 27 0 36 1694 1738 5 5 3506
15760 (java) 1884 1607 8 9 0 11 0 2 3521
15761 (java) 3 26 2 10 0 2 1553 1929 3525

16536 (java) 2 0 1 23 1 31 1934 1526 3517
16537 (java) 1831 1614 0 9 0 33 2 11 3501
16538 (java) 1 0 0 5 1940 1537 2 11 3497
16539 (java) 1 1 1596 1868 0 25 12 8 3511

17504 (java) 1 1 1 9 1769 1687 3 45 3516 <
17505 (java) 1 4 1 6 6 3 1522 1971 3514
17506 (java) 1413 1 9 2039 0 9 2 33 3506 <
17507 (java) 114 1887 1430 5 7 16 3 37 3500 <

17933 (java) 0 1 0 14 1707 1751 8 28 3510
17934 (java) 0 0 1745 1729 0 7 7 33 3522
17935 (java) 1700 1771 4 25 10 3 1 28 3542
17936 (java) 16 15 0 9 0 0 1702 1765 3507

18344 (java) 5 0 7 4 1567 1910 2 11 3507
18345 (java) 1 3 1557 1932 17 26 6 8 3550
18346 (java) 8 0 2 10 4 29 1317 2124 3497
18347 (java) 1591 1856 0 8 8 28 0 16 3508

18753 (java) 1737 1786 4 5 0 1 0 18 3552
18754 (java) 5 28 0 4 1617 1833 3 26 3515
18755 (java) 21 27 0 6 0 0 1518 1937 3509
18756 (java) 1 25 1690 1785 0 0 0 8 3509

19166 (java) 41 1 1829 1628 0 9 3 13 3523
19167 (java) 1737 1736 3 8 0 1 10 15 3511
19168 (java) 33 8 8 13 2 2 1546 1905 3517
19169 (java) 36 9 4 5 1634 1802 0 10 3500

19983 (java) 1617 1818 8 34 12 5 3 2 3501
19984 (java) 0 1 2 41 0 0 1724 1743 3513
19985 (java) 1 3 1901 1570 0 0 3 9 3487
19986 (java) 3 21 0 32 1634 1816 0 6 3514

20947 (java) 0 0 2 27 1652 1819 0 3 3506
20948 (java) 0 0 2 9 1 33 1848 1635 3528
20949 (java) 1 0 1873 1597 1 31 0 4 3508
20950 (java) 1624 1821 10 9 3 33 0 3 3503

21354 (java) 1 17 30 8 601 1400 1 1445 3503 <
21355 (java) 0 3 1542 1963 0 16 1 9 3536
21356 (java) 1761 1672 27 8 8 8 1 10 3494 <
21357 (java) 0 0 25 3 1221 8 1843 410 3511

1b) 2 x 4 socket wide instances (each instance has 40 warehouse threads)

Note: The correct NUMA nodes got picked 9 out of 10 runs.

PID Node 0 Node 1 Node 2 Node 3 Node 4 Node 5 Node 6 Node 7 Total
21847 (java) 1 0 1 16 821 1154 1241 1330 4563
21848 (java) 1274 1497 917 807 14 3 1 25 4537

22120 (java) 1 26 1 5 860 948 1367 1348 4555
22121 (java) 1269 1392 929 896 1 1 5 13 4505

23298 (java) 847 829 1346 1470 0 30 7 3 4533
23299 (java) 3 4 0 15 876 1130 1184 1308 4521

23603 (java) 1277 1447 951 817 2 1 8 20 4521
23604 (java) 1 10 1 28 876 831 1343 1460 4551

23874 (java) 0 3 2 37 864 853 1309 1473 4542
23875 (java) 1452 1382 859 873 1 1 1 10 4579

24148 (java) 1 1 1 21 1132 1296 1084 1039 4574
24149 (java) 896 836 1223 1615 0 30 1 8 4611

24421 (java) 1447 1413 818 849 1 0 0 24 4551
24422 (java) 1 27 1 6 865 817 1420 1427 4563

25064 (java) 861 907 1344 1404 0 24 0 7 4547
25065 (java) 1 1 0 14 1197 1376 827 1168 4583

25881 (java) 1 3 0 19 1043 1204 1123 1167 4559
25882 (java) 884 874 1402 1323 0 0 1 37 4522

26150 (java) 1 1 6 781 11 886 1347 1514 4546 <
26151 (java) 1129 1322 811 459 760 11 9 29 4531 <


----
2) 8 socket IvyBridge-EX prototype system with the following SLIT table :

node distances:
node 0 1 2 3 4 5 6 7
0: 10 16 30 30 30 30 30 30
1: 16 10 30 30 30 30 30 30
2: 30 30 10 16 30 30 30 30
3: 30 30 16 10 30 30 30 30
4: 30 30 30 30 10 16 30 30
5: 30 30 30 30 16 10 30 30
6: 30 30 30 30 30 30 10 16
7: 30 30 30 30 30 30 16 10

2a) 4 x 2 socket wide instances (each instance has 60 warehouse threads)

PID Node 0 Node 1 Node 2 Node 3 Node 4 Node 5 Node 6 Node 7 Total
18680 (java) 2 18 1 20 1 12 3626 4820 8500
18681 (java) 1 8 5 174 4721 3579 12 3 8504
18682 (java) 4 18 5002 3461 9 13 13 3 8523
18683 (java) 5045 3428 3 7 1 18 6 3 8511

19745 (java) 5472 2984 4 7 8 22 10 20 8527
19746 (java) 8 30 3467 4976 3 22 7 5 8519
19747 (java) 2 18 4 9 4390 4068 1 2 8493
19748 (java) 5 7 2 11 4 16 4251 4186 8482

20796 (java) 4176 4301 4 6 9 13 5 6 8517
20797 (java) 13 47 27 1 3311 5127 5 3 8535
20798 (java) 6 5 2973 5498 3 19 29 11 8544
20799 (java) 9 7 9 130 2 14 3034 5315 8519

21852 (java) 6 1 14 14 24 2043 2346 4035 8484
21853 (java) 3551 4841 10 23 4 11 86 4 8532
21854 (java) 2 1 3194 5286 2 5 17 12 8519
21855 (java) 11 3 1 12 6943 1288 166 78 8502

22907 (java) 5 1 4165 4307 8 15 21 8 8531
22908 (java) 4 10 2 2 3066 5421 4 1 8511
22909 (java) 3 2 5 8 6 16 3544 4943 8527
22910 (java) 3340 5127 1 18 6 16 9 4 8520

23955 (java) 7 2 19 3 2436 6000 2 47 8515
23956 (java) 12 2 2785 5668 25 5 3 25 8525
23957 (java) 3651 4790 15 2 24 8 1 17 8508
23958 (java) 9 4 5 1 13 5 4794 3688 8519

25011 (java) 3532 4931 1 6 8 7 3 32 8520
25012 (java) 7 14 2 8 2763 5683 1 28 8506
25013 (java) 13 12 1 11 18 5 5373 3095 8528
25014 (java) 15 9 5455 3013 4 7 8 11 8522

26060 (java) 1 1 1 3 29 28 4651 3784 8498
26061 (java) 2 1 6245 2212 13 39 0 5 8519
26062 (java) 68 5294 1 3 1879 1262 1 3 8511
26063 (java) 7522 637 1 78 107 160 1 6 8512

27116 (java) 2 37 2 2 9 5 5110 3368 8535
27117 (java) 5376 3115 12 4 10 3 5 6 8532
27118 (java) 9 22 2850 5610 4 17 1 18 8531
27119 (java) 10 26 2 3 3472 4987 1 5 8505

28162 (java) 3332 5103 1 4 8 4 7 45 8504
28163 (java) 12 4 3 18 3079 5383 1 24 8524
28164 (java) 17 9 5533 2921 9 5 1 16 8510
28165 (java) 7 2 1 16 10 7 5166 3299 8508


2b) 2 x 4 socket wide instances (each instance has 120 warehouse threads)

PID Node 0 Node 1 Node 2 Node 3 Node 4 Node 5 Node 6 Node 7 Total
29228 (java) 2 8 2394 3868 15 82 1130 1085 8585
29229 (java) 2824 3773 19 6 963 923 24 59 8589

29964 (java) 1084 1080 28 31 2841 3497 8 2 8569
29965 (java) 16 18 1186 1077 3 6 2841 3436 8583

30683 (java) 19 4281 1830 2137 73 94 13 26 8472
30684 (java) 3763 31 10 10 75 56 2174 2394 8514

31399 (java) 6 25 2608 3116 10 18 1424 1388 8595
31400 (java) 1518 1532 2 12 2259 3188 20 32 8562

32113 (java) 692 620 28 18 3141 4033 25 20 8577
32114 (java) 8 8 2231 2390 3 15 1955 1974 8583

32827 (java) 2324 2663 79 3423 24 20 8 13 8554
32828 (java) 15 3 2448 26 515 855 2214 2489 8566

33543 (java) 1090 1129 3 23 22 25 2923 3366 8581
33544 (java) 12 10 2686 3375 1745 726 4 22 8580

34258 (java) 2127 1928 1924 2558 7 18 8 4 8574
34259 (java) 5 23 8 11 1182 1266 2731 3347 8572

34978 (java) 16 14 1617 1964 14 23 2258 2677 8583
34979 (java) 1721 1704 62 19 2339 2726 6 10 8588

35692 (java) 3599 2721 50 58 6 9 700 1421 8564
35693 (java) 18 49 74 83 3162 4822 309 46 8563



\
 
 \ /
  Last update: 2014-10-10 21:21    [W:0.493 / U:0.052 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site