Threading Performance

Paderborn Center for Parallel Computing: Jens Simon (simon(at)upb.de)

We are measuring different kinds of performance data of shared memory multiprocessor systems. At first, the synchronization time of the programming methods OpenMP, POSIX Threads, and Message Passing Interface are considered depending on the number of synchronized threads. Also the accumulated memory bandwidth of different numbers of threads performing the Triad-operation is presented.

	Fujitsu Siemens TX300-S3				Supermicro X7DB8+			Supermicro H8QMi-2				Fujitsu Siemens V830			Tyan S4882			Supermicro X6DAE-G2		Sony Playstation 3
	2-socket Quad-Core Intel Xeon 5355, 2.66 GHz, Intel 5000X, FSB1333, 4 x 2 GByte DDR2-667				2-socket Dual-Core Intel Xeon 5160, 3.0 GHz, Intel 5000V, FSB1333, 4 x 2 GByte DDR2-555			4-socket Dual-Core AMD Opteron 8220 2.8 GHz, HT1000, nVidia MCP55 Pro, 32x2 GByte DDR2-555				2-socket Dual-Core AMD Opteron 270, 2.0 GHz, nVidea2200, HT1000, 2 x (2x2) GByte DDR1-333			4-socket Singel-Core AMD Opteron 848, 2.2 GHz, AMD 8131, HT1000, 4 x (4x2) GByte DDR1-333			2-socket Singel-Core Intel Xeon DP, 3.2 GHz, Intel E7525, FSB1066, 8 x 0,5 GByte DDR2-400		Cell Broadband Engine 3.2 GHz, 256 MByte XDR, Hypervisor 6 SPEs, 200 Mbyte


Total Number of Cores	1	2	4	8	1	2	4	1	2	4	8	1	2	4	1	2	4	1	2	1	2	4	6
OpenMP Barrier [us]	0,35				0,31			0,19				0,25			0,23			0,47
		1,89				1,78			1,37				1,52			1,78			1,90
			2,45				2,29			2,48				2,42			2,63
				3,92							3,73
PThread Barrier [us]	0,30				0,27			0,35				0,25			0,24			0,55
		8,1				12,3			7,8				8,9			10,1			8,2
			17				18			19				21			24
				37							38
MPI Barrier [us]	-				-			-				-			-			-		-
		1,01				0,98			0,52				0,96			1,10			0,99
			1,76				1,43			1,19				1,74			1,84
				3,74							2,4
OpenMP TRIAD [MB/s]	3139				3190			4400				2475			2169			2838		2100
		4918				5222			8500				4386			4249			2469		4000
			5076				5305			16100				4985			5768					7500
				5083							18000												10200

More details of the measurements can be found here: