Mindcraft NT vs. Solaris SMB performance tests

[Kalle Kiviaho]

Andrew Miller wrote:

> Has anyone tried to reproduce the NT/Solaris SMB server comparison done
> by Mindcraft:

>         http://www.mindcraft.com/whitepapers/nt4sol26exec.html
 
I've run some performance tests on our Sun E450 and Sun E4000. 
The tests where done on Samba 1.9.18p8. Samba was compiled with
FAST_SHARE_MODES and MMAP. In the smb.conf the "socket options = 
TCP_NODELAY" was set.

--- test start ---

E450:	Model:		Sun Enterprise 450
	Processors:	2 * 250 UltraSparc II
	Memory:		512MB
	Network:	100Mbps

E4000:	Model:		Sun Enterprise 4000
	Processors:	4 * 250 UltraSparc II
	Memory:		1GB
	Network:	2 * 100Mbps, 2 * 400Mbps (QuadFastEthernet /w
			SunTrunk software). The test was done on a
			400Mbps trunk.

The tests where done at 2 different times of the day and where repeated
3 days in a row:

	During working hours:	The E450 has about 150 users connected
				to it during working hours.
				The E4000 has about 400 users connected
				to it during working hours. ~200 of them
				through the used networkinterface.

	During the night:	In this part the system did not have
				any other samba users. This was done
				with 10 PC clients (setup: P200MMX,
				64MB, 2GB-IDE disk, 100Mbps network)
				copying ~500MB of random files from 
				different shares from the server.
					
The E450 handles DNS and mail for our network, the load for this is
almost negliable.

				E450

			PEAK		AVG		LOW
Working hours:	CPU:	27%		2%		1%
		NET:	8.5MB/s		1.5MB/s		0.0MB/s
		IOW:	6%		0%		0%

Night:		CPU:	6%		2%		1%
		NET:	9.0MB/s		7.5MB/s		0.5MB/s
		IOW:	3%		0%		0%

				E4000

			PEAK		AVG		LOW
Working hours:	CPU:	22%		2%		<1%
		NET:	32.5MB/s	5.5MB/s		0.0MB/s
		IOW:	6%		0%		0%

Night:		CPU:	3%		1%		<1%
		NET:	35.0MB/s	26.0MB/s	0.5MB/s
		IOW:	3%		0%		0%

CPU	= 	CPU utilization
NET	= 	Network utilization
IOW	=	IO Wait
	
The "high" CPU peaks during working hours seem to be when lot of
people print at the same time. The network util stats where gathered
from the Cisco 5500 switch that serves our server room and distributes
the data to "sub"-switches on the office floors. 

--- test end ---

The BIG magic to benchmarking is FAST disks, that's what we have on
our systems (most of the data on the E4000 comes from a SparcStorage
array).

I really hope they would have disclosed the setup of the test machines.

/mvh Kalle Kiviaho - kivi at ios.chalmers.se

PS. I'm going to run these test with NetBench in the near future.. :) DS.