US QCD Machine Performance

USQCD Machine Performance

Machine	Processor per node	total no. of nodes	total no. of cores	DWF per node	Clover per node	asqtad per node	Jpsi Equivalence
kaon	2.0 GHz Dual CPU Dual Core Opteron	600	2400	4696 MFlops	3180 MFlops	3832 MFlops	0.88 Jpsi-core-hour
jpsi	2.1 GHz Dual CPU Quad Core Opteron	856	6848	10061 MFlops	7423 MFlops	9563 MFlops	1 Jpsi-core-hour
7n	1.9 GHz Dual CPU Quad Core Opteron	396	3168	8800 MFlops	5148 MFlops	6300 MFlops	0.77 Jpsi-core-hour
9q	2.4 GHz Dual CPU Quad Core Nehalem	320	2560	MFlops	MFlops	MFlops	1.94 Jpsi-core-hour
9g	GPU	200	-	MFlops	MFlops	MFlops	TBD
QCDOC	400 MHz PPC Core estimated	12288	12288	336 MFlops	-	360 MFlops	0.24 Jpsi-core-hour
BlueGene/P	850 MHz Quad Core PowerPC 850	1024 per rack	4096 per rack	2560 MFlops	2511 MFlops	2680 MFlops	0.54 Jpsi-core-hour
Cray XT5	2.1 GHz Quad Core Opteron	7832	31328	-	2232 MFlops	2260 MFlops	0.50 Jpsi-core-hour

WARNING:

Performance numbers for the 9q cluster are preliminary estimates and will be updated soon. The conversion factors for the 9g GPU strongly depend on the usage model and will be determined at a later stage.

COMMENTS:

The table above shows the measured performance of DWF, anisotropic clover, and asqtad inverters on the kaon, jpsi, 7n and 9q clusters, on the ANL BG/P, the ORNL XT5 and the QCDOC. All performance numbers are single precision unless otherwise noted.

The DWF, Clover and asqtad performance figures for kaon, jpsi and 7n used 128-process (32-node, 16-node, 64-node,and 16-node respectively) runs, with 4, 2, or 8 processes per node, one process per core. DWF and Clover data were taken with Chroma. kaon and jpsi Clover runs used 6^3x64 local (per core) lattices, and DWF runs used 14x7x7x16 local (per core) lattices with Ls=16. Clover performance on 7n used 4^3x8 local volumes per process, DWF performance used global volume 24^3x64 (Ls=16), and asqtad performance used local volume 14^4.

The QCDOC DWF (double precision) and asqtad (single precision) estimates are based on the observed peak performance of the double precision conjugate gradient codes on early motherboards, scaled to 400 MHz. Clover performance data are not available.

The BG/P asqtad result is the average of the performance of 6^4 and 8^4 local volumes, and is single precision. The DWF result is double precision, using 4^4 (Ls=16) local volumes. The Clover result used 4096 cores.

The XT5 Clover performance figure is based on anisotropic Clover calculations on 32^3x256 global volume run on 24 cores (Robert Edwards) and HISQ runs on 64^3x128 lattices on 2k cores (Steve Gottlieb).

The final column of the table gives the Jpsi-equivalence for each of the USQCD resources. All except the Cray XT5 use the ratio of the average performance of asqtad and DWF; the XT5 uses the ratio of the average performance of the asqtad (HISQ) and clover inverters. Also, the QCDOC Jpsi-equivalence figure of 0.24 has been assigned to be consistent with prior years' accounting, rather than using the estimated DWF and asqtad performance values.