Talk:NAMD Benchmarks

The efficiency of a parallel system describes the fraction of the time that is being used by the processors for a given computation. It is defined as

        Execution time using one processor       ts
E(n)= --------------------------------------  = ----
      N * Execution time using N processors     N tn

In general, parallel jobs should scale to at least 70% efficiency. The ASC's DMC recommends a scaling efficiency of 75% or greater. For NAMD the efficiency of a parallel job can be calculated as follows (where N is processors committed to the job):

  days/ns where N = 1
 --------------------- * 100 = Efficiency
      N * days/ns

Information on NAMD performance Scaling is available at: http://www.ks.uiuc.edu/Research/namd/wiki/?NamdPerformanceTuning

Benchmark used for performance evaluation on the Cheaha, Biowulf, and DMC is Apoa1 from the NAMD suite and is available at: http://www.ks.uiuc.edu/Research/namd/utilities/

The parameters for the benchmark used throughout are: 500 steps, 92K atoms, 12A cutoff + PME every 4 steps.

Sample Benchmark using NAMD

Sample Benchmark comparing Days/ns for Cheaha, Biowulf, and DMC using InfiniBand

Days/ns (efficiency) on Cheaha, Biowulf, and DMC

Processors	Cheaha	Biowulf	DMC <ref group="note" >No hardware control was possible for the ASC's DMC and as such it is possible that different runs might have run on different hardware. *Efficiency on DMC using 4 processors is above 100%. This could be probably because of an older generation processor being used to run the job serially i.e. using 1 processor, thus setting a lower baseline to begin with causing the efficiency to go up as better processors are available for the parallel jobs.</ref>
1	15.4054 (100%)	18.0535 (100%)	19.1000 (100%)
2	7.7119 (99.87%)	9.5163 (94.86%)	9.7600 (97.84%)
4	3.8933 (98.92%)	4.9222 (91.69%)	4.7570 (100.3%)*
8	1.9653 (97.98%)	2.5763 (87.59%)	2.5360 (94.14%)
16	0.9950 (96.76%)	1.2658 (89.14%)	1.3870 (86.06%)
32	0.5101 (94.37%)	0.6463 (87.29%)	0.7438 (80.24%)
64	0.2592 (92.83%)	0.3390 (83.22%)	0.3938 (75.78%)
128	0.1360 (88.45%)	NA	NA
256	0.0770 (78.09%)	NA	NA
512	0.0439 (68.49%)	NA	NA

Bench mark notes

The above benchmarks were run using NAMD 2.8b1 and the Apoa1 benchmark suite from NAMD.

On Cheaha, only the third generation hardware was used for the above benchmarks. More information about the hardware used on Cheaha is available here.

The data for the NIH-Biowulf benchmarks is available at: http://biowulf.nih.gov/apps/namd/namd_bench.html. The Biowulf hardware used for comparison in the e2800 with Infiniband.

ASC-DMC

<references group="note"/>

Comparison of Ethernet and IB

Ethernet vs Infiniband interconnect on Cheaha Gen 3 hardware

Processors	Infiniband	Ethernet
1	15.4054 (100%)	15.4040 (100%)
2	7.7119 (99.87%)	7.7800 (98.99%)
4	3.8933 (98.92%)	3.9405 (97.72%)
8	1.9653 (97.98%)	2.2816 (84.39%)
16	0.9950 (96.76%)	1.2714 (75.74%)
32	0.5101 (94.37%)	0.6973 (69.03%)
64	0.2592 (92.83%)	0.6562 (36.67%)
128	0.1360 (88.45%)	0.8950 (13.44%)
256	0.0770 (78.09%)	0.8632 (6.97%)
512	0.0439 (68.49%)	NA

Comparison on different Generations of Cheaha hardware on NAMD with Infinband

Cheaha Gen 2 and Gen 3 hardware using NAMD 2.8 and Inifiband

Processors	Generation 3	Generation 2
1	15.4054 (100%)	15.9300 (100%)
2	7.7119 (99.87%)	8.6620 (91.98%)
4	3.8933 (98.92%)	4.2560 (93.59%)
8	1.9653 (97.98%)	2.3450 (84.93%)
16	0.9950 (96.76%)	1.2060 (82.56%)
32	0.5101 (94.37%)	0.6420 (77.57%)
64	0.2592 (92.83%)	0.2750 (90.38%)
128	0.1360 (88.45%)	0.1620 (76.81%)
256	0.0770 (78.09%)	NA
512	0.0439 (68.49%)	NA

Hardware Notes

Generation 3 hardware contains 576 cores. 2.66 GHz Intel Xeon Hex-core (48-2x6 core) Generation 2 hardware contains 192 cores. 3.06 GHz Intel Xeon Quad-core (24-2x4 core) More information about the hardware used on Cheaha is available here.

Actual job benchmarks (Segrest job)

Actual NAMD job courtesy of Center for Computational and Structural Biology.

STRUCTURE SUMMARY:

Job parameters: 100,000 steps, 246K atoms, 12A cutoff + PME every 4 steps
245956 ATOMS
174192 BONDS
130742 ANGLES
84410 DIHEDRALS
1622 IMPROPERS
0 CROSSTERMS
0 EXCLUSIONS
232802 RIGID BONDS
505063 DEGREES OF FREEDOM
84698 HYDROGEN GROUPS
4 ATOMS IN LARGEST HYDROGEN GROUP
84698 MIGRATION GROUPS
4 ATOMS IN LARGEST MIGRATION GROUP
TOTAL MASS = 1.48179e+06 amu
TOTAL CHARGE = 2.7746e-05 e
MASS DENSITY = 1.01378 g/cm^3
ATOM DENSITY = 0.101334 atoms/A^3

Speedup (Actual Wall times, days/ns, efficiency)

Actual NAMD CSB job on Cheaha Gen 3 hardware

Processors	Wall Clock	Days/ns	Speedup
1	357278 (99hrs 14mins)	20.4950 (100%)	1
8	45784 (12hrs 42mins)	2.6237 (97.64%)	7.80
16	23380 (6hrs 29mins)	1.33205 (96.16%)	15.28
32	11871 (3hrs 17mins)	0.6809 (94.04%)	30.09
64	6242 (1hr 43mins)	0.3626 (88.30%)	57.23
128	3386 (56mins 26secs)	0.1940 (82.51%)	105.51
256	1814 (30mins 14secs)	0.0982 (81.51%)	196.96
512	1129 (18mins 49secs)	0.0578 (69.21%)	316.46

Actual job comparison to IBM BG/L

NAMD Comparison data to UAB IBM BlueGene/L.

Job parameters: 100,000 steps, 246K atoms, 12A cutoff + PME every 4 steps

NAMD Comparison to IBM BG/L

System	CPU count	Atoms	Days/ns
UAB BlueGene/L	256	246K	0.80
Cheaha Gen3	32	246K	0.6809
Cheaha Gen3	256	246K	0.0982
Cheaha Gen3	512	246K	0.0578

The Cheaha outperforms the BG/L using only 32 cores on the gen 3 hardware. At equal processor count the Cheaha is roughly 8 times faster than the BG/L.