Single-core Intel Xeon 2.8GHz on FreeBSD 6.1: SMP+HTT Testing

The following data is the result of a series of tests comparing performance of an SMP kernel with HyperThreading enabled (machdep.hyperthreading_allowed=1) versus a non-SMP kernel resembling GENERIC (minus built-in support for a few cards) on a single-core Xeon.

The test box is a Dell PowerEdge SC1425 with 1GB of RAM with an Intel Xeon 2.8GHz processor running FreeBSD 6.1-RELEASE-p10.

The preformatted text before each table gives an idea of the command that generated the data. The tests (other than buildkernel) were done using sysbench 0.4.8. I've never used the program before so I can't vouch for its accuracy.

dmesg: ALVIS-SMP or ALVIS

make -j buildkernel

# rm -rf /usr/obj/usr/src/sys/
# time env -i make -jXX buildkernel KERNCONF=ALVIS

	real	user	sys
-j2 (SMP)	11:52.60	18:20.81	01:51.07
-j2	13:31.32	11:45.43	01:30.69
-j4 (SMP)	11:43.61	18:07.96	01:56.37
-j4	13:31.95	11:50.68	01:32.21

Here we can see that the least real time was taken when the build was done with 4 jobs and an SMP+HTT setup. The best SMP case (-j4) shows a decrease of about 18.6% real build time from the best non-SMP (-j2) case.

sysbench cpu test

# sysbench --test=cpu --num-threads=XX run

		Total Event Time (s)
	Threads	Normal	SMP+HTT	Change
TRIAL 1	1	55.1678	55.1733	0%
	5	275.7409	218.264	-20.80%
	10	556.5996	435.8256	-21.7%

TRIAL 2	1	55.2194	55.1258	-0.10%
	5	275.8694	218.2093	-20.90%
	10	551.7829	436.0049	-21.00%

Here we see the benefit of SMP+HTT when testing the CPU using the sysbench tool (benchmarks/sysbench) with multiple threads. I still need to try even threads (2,4,8)... I guess I wasn't thinking.

sysbench memory write test

# sysbench --test=memory --memory-total-size=10G --num-threads=XX run

		Total Event Time (s)		Operations/sec		Rate (MB/s)
	Threads	Normal	SMP+HTT	Normal	SMP+HTT	Normal	SMP+HTT
TRIAL 1	1	51.5733	51.9178	147992.47	146781.08	144.52	143.34
	2	104.3083	70.4953	147453.09	218100.46	144	212.99
	4	206.3408	137.1053	145875.86	219819.52	142.46	214.67
	8	392.3044	258.2406	147359.85	219601.62	143.91	214.45

TRIAL 2	1	51.0001	52.0631	148204.53	145734.67	144.73	142.32
	2	104.0735	69.9713	147436.35	220046.93	143.98	214.89
	4	208.3495	137.1801	146383.1	219867.61	142.95	214.71
	8	394.3918	256.6188	147236.78	219649.53	143.79	214.5

The above test by sysbench is a test of writing (default) 10GB (default=100GB) to sequential memory (default) with memory block sizes of 1K (default).

It is clear that with multiple threads, the HyperThreading/SMP wins out by a lot, with rate increases of up to 50%! However, when the application uses a single thread, there is a measurable loss using SMP+HTT, and the non-SMP kernel performs best (but not by much).

The graphs below were generated using data from TRIAL 1.