NUMA schedulers tests

Hi,

here is a set of measurements for three benchmarks and five schedulers. They were performed on an NEC Azusa, 16 CPU Itanium machine with NUMA architecture, 4 CPUs per node, 4 nodes with 8GB memory each.

The benchmarks are:
- hackbench: a variant of the wellknown chatroom benchmark
- numa_test: runs a certain number of memory bandwidth hungry and latency dependent processes in parallel
- kernbench: parallel kernel compile (5 times, make -j24, compile an ia64 2.4.18 kernel with gcc 2.96. As gcc's speed is VERY dependent on the number of registers of the target architecture, the results are not comparable to i386 compiles.)

Schedulers tested (all on top of discontigmem for ia64):

A: vanilla O(1) scheduler from 2.5.39
B: Michael Hohnbaums simple NUMA scheduler (latest published rev 2)
C: pooling NUMA scheduler with initial load balancing (patches 01+02
from the 5 patches sent out)
D: node affine NUMA scheduler (patches 01+02+03)
E: node affine NUMA scheduler with dynamic homenode (patches 01+02+03+05)

All results are averages over several measurements, the numbers in braces are the standard deviation or "error bars".

There's a lot to say about the results, e.g that E is the best because it has most features. But the only comment I really want to make on this now
is: the numa_test is not good enough for testing node affinity as "hackbench 20" doesn't seem to disturb the initialy balance too much. I'll improve this to make it more realistic. Otherwise it's good enough to see the advantage of not moving around tasks across the nodes.

Best regards,
Erich

Hackbench: (averages over 4 tests, execution time)
-------------------------------------------------------------------
N: 10 25 50 100
A 5.54(0.10) 15.28(0.03) 32.65(0.35) 67.67(0.69)
B 5.37(0.06) 15.13(0.44) 31.96(0.49) 67.14(0.43)
C 1.31(0.08) 2.91(0.12) 5.95(0.13) 13.37(0.11)
D 2.45(0.44) 8.89(1.47) 21.98(1.08) 30.57(17.10)
E 1.39(0.08) 2.92(0.09) 6.20(0.34) 14.30(0.28)

numa_test: N=4
----------------------------------------------------------------------------
Elapsed AvgUserTime TotUserTime TotSysTime Runs
A 42.95(3.23) 30.40(1.16) 121.65(4.62) 10
B 35.08(2.24) 28.51(0.77) 114.10(3.07) 0.19(0.03) 20
C 27.44(0.06) 27.14(0.04) 108.62(0.15) 0.16(0.01) 20
D 27.42(0.08) 27.10(0.04) 108.48(0.14) 0.15(0.01) 20
E 27.50(0.23) 27.10(0.03) 108.48(0.13) 0.15(0.01) 20

numa_test: N=8
----------------------------------------------------------------------------
Elapsed AvgUserTime TotUserTime TotSysTime Runs
A 42.75(2.31) 31.34(0.93) 250.82(7.46) 10
B 34.82(2.74) 29.76(0.67) 238.19(5.34) 0.44(0.04) 20
C 29.29(0.47) 28.80(0.14) 230.55(1.13) 0.35(0.02) 20
D 29.23(0.40) 28.76(0.12) 230.20(0.92) 0.34(0.02) 20
E 29.08(0.06) 28.76(0.03) 230.22(0.27) 0.34(0.01) 20

numa_test: N=16
----------------------------------------------------------------------------
Elapsed AvgUserTime TotUserTime TotSysTime Runs
A 47.02(0.88) 33.80(0.90) 541.15(14.47) 10
B 41.82(8.26) 32.04(0.47) 512.91(7.57) 1.08(0.05) 20
C 41.30(5.50) 31.80(0.18) 509.11(2.84) 1.00(0.05) 20
D 39.89(6.42) 31.80(0.10) 509.13(1.67) 0.95(0.03) 20
E 41.61(5.42) 31.76(0.27) 508.44(4.32) 0.96(0.04) 20

numa_test: N=32
----------------------------------------------------------------------------
Elapsed AvgUserTime TotUserTime TotSysTime Runs
A 83.52(2.25) 37.44(1.16) 1198.57(37.18) 10
B 80.33(4.35) 33.86(0.60) 1083.94(19.07) 2.13(0.06) 10
C 77.84(4.10) 33.58(0.36) 1074.82(11.47) 4
D 73.59(6.31) 33.20(0.11) 1062.88(3.38) 4
E 69.86(0.41) 33.26(0.06) 1064.72(1.88) 2.01(0.03) 10

numa_test: N=64
----------------------------------------------------------------------------
Elapsed AvgUserTime TotUserTime TotSysTime Runs
A 164.15(3.40) 38.73(0.77) 2479.61(49.41) 10
B 149.63(2.89) 34.61(0.40) 2215.80(25.42) 4.28(0.06) 10
C 147.22(1.82) 34.41(0.13) 2202.88(8.39) 4
D 139.32(5.26) 33.37(0.05) 2136.66(3.59) 4
E 137.10(2.15) 33.40(0.02) 2138.13(1.37) 4.08(0.05) 10

Kernbench: (averages over 5 compiles)
-----------------------------------------------
Elapsed UserTime SysTime
A 93.78(0.43) 1321.05(1.20) 49.31(0.14)
B 93.47(1.57) 1303.11(0.53) 52.27(0.46)
C 92.67(1.08) 1304.09(0.74) 52.79(0.32)
D 93.19(1.45) 1300.30(0.89) 52.46(0.20)
E 93.63(1.00) 1299.44(0.67) 52.81(0.19)

<snip>

Thanks very much for doing this Erich. I have run the SPEC SDET benchmark on 2.5.41-mm3, with your "E" scheduler, and Michael's rev 1 & 2 scheduler. At this time I don't think I can publish actual throughput results, but only % differences, to abide by SPEC rules. These results are not considered compliant by SPEC, and are used only to compare performance difference of these scheduler implementations.

SDET simulates the multiuser development environment, and may be useful for
environments like universities, etc. More information can be found at:
http://www.spec.org/osg/sdm91/

SDET was configured to simulate 1, 2, 4, 8, 16, 32, 64 and 128 users. SDET appears to benefit from spreading out exec'd processes on initial placement, but there also exists scenarios where exec'd tasks should reside on the same node. Ideally, each of the scripts run in parallel would benefit from being initially spread out across the nodes, but the processes exec'd _within_ each script would benefit from being placed within the node. I don't expect any scheduler to to identify and optimize for that today, but maybe that's something to work on in the future, who knows?

Anyway, on to the results. Almost forgot, this is on a 16 way NUMA-Q box. These results are not actual results, but normalized to a baseline for 2.5.41-mm3 for 1 user:

1 user
vanilla 100
erich-E 062
michael-r1 103
michael-r2 065

2 users
vanilla 191
erich-E 112
michael-r1 180
michael-r2 116

4 users
vanilla 288
erich-E 213
michael-r1 306
michael-r2 208

8 users
vanilla 426
erich-E 336
michael-r1 405
michael-r2 333

16 users
vanilla 462
erich-E 412
michael-r1 464
michael-r2 427

32 users
vanilla 438
erich-E 425
michael-r1 436
michael-r2 332

64 users
vanilla 420
erich-E 410
michael-r1 430
michael-r2 432

128 users
vanilla 413
erich-E 397
michael-r1 413
michael-r2 398

In Michael's rev 1, initial placement of an exec'd task would be on the same CPU if nr_running <=2, (sched_best_cpu) which I believe why the low loaded tests perform better. This is where I believe the tasks within a script are getting scheduled on the same CPU/node and benefit from locality. Erich's and Michael's rev2 place the newly exec'd task on the least loaded cpu, which tends to spread those tasks exec'd within a script across nodes and hurt thoughput in those cases. In the higher load cases, most of the results appear similar.

Anyway, I used SDET to give yet another view of how these schedulers can affect performance. At this point I personally do not have a preference on which scheduler one to use. I'd like to hear from people who have made BIG changes to the scheduler (hint, hint) what they think of these implementations, and where we should go from here. We should probably decide which implementation to use, and what we need to do to make it ready for inclusion, right?

Andrew Theurer

Erich,

Thanks for the numbers. These are quite interesting. Your scheduler which goes through much effort to spread the load evenly across all of the nodes appears to do much better on small loads, but as the system becomes loaded, the difference gets smaller. Still, your best combination is the clear winner on the numa_test. But on kernbench, that changes and all the runs look quite close.

While numa_test is interesting and might be quite relevant to the HPC loads that you focus on, I'm not sure it should be the defining test for NUMA scheduler performance. The main problem is that we have differing workloads that we care about.

I would be happy to see you NUMA scheduler get accepted into mainline, as I think there are enough knobs that it can be tuned for varying workloads. It will just require some time and testing to establish these.

On the other hand, my NUMA scheduler is a definite improvement for NUMA boxes, so if your scheduler cannot be included in mainline, getting mine in would be of benefit.

Ingo, do you have any thoughts on this you would be willing to pass along?

Michael

Michael Hohnbaum                      503-578-5486
[email protected]                   T/L 775-5486