Mac OS X SAMBA 2.0.3 time test

Wed Jun 23 20:47:18 GMT 1999

I'm not sure if this is the right Samba list to send this data to, but I
thought it might be interesting for samba people to look at.  Note the
brain dead performance with Windows 98 and Samba 2.0.3 on a 100MB pipe. :(

--snip--

SMB/CIFS time trials
-------------------

Goal:  To do file transfer time tests between NT servers native SMB/CIFS
services and 
that of Mac OSX Samba SMB/CIFS services.  Estimated results is that Samba
is ~25% 
slower than NT server native.  Using NT and 98 clients.

-------------------

Machine A
	DELL 4300
	P2-400
	128MB RAM
	9GB RAID 5 HD 32MB cache
	SCSI
	NT SRV 4.0 SP5

Machine B
	Mac G3
	PPC 750 400MHZ
	256MB RAM
	9GB HD
	SCSI
	MAC OSX 1.0.1
	SAMBA 2.0.3

Machine C
	Homemade
	P-100
	32MB RAM
	18GB HD
	SCSI
	NT SRV 4.0 SP4

Machine D
	IBM Thinkpad 600
	P2-266
	96MB
	4GB HD
	ATA IDE
	Windows 98

-------------------

Package 1
	14 files
	~10MB each
	126MB total

Package 2
	15 files
	~5MB each
	60.2MB total

Package 3
	31 files
	~1MB each
	41.8MB total

Package 4
	47 files
	~4MB each
	214MB total

-------------------

Lingo:

X write Y to Z (from machine X paste directory containing package Y to
machine Z)
X read Y from Z (from machine X copy directory containing package Y from
machine Z)

-------------------

Private 10MB shared ethernet subnet	Private 100MB shared ethernet subnet

[ client writes to server ]		[ client writes to server ]

C write 1 to A - 130 sec		C write 1 to A -  65 sec (~25% net util)
C write 1 to B - 130 sec		C write 1 to B -  27 sec (~35% net util)

C write 2 to B -  60 sec		C write 2 to B -  14 sec (same %s)
C write 2 to A -  60 sec		C write 2 to A -  40 sec

C write 3 to A -  45 sec		C write 3 to A -  20 sec
C write 3 to B -  45 sec		C write 3 to B -  11 sec

[ client reads from server ]		[ client reads from server ]

C read 3 from B - 43 sec		C read 3 from B - 12 sec (~35% net util)
C read 3 from A - 41 sec		C read 3 from A - 12 sec (~35% net util)

C read 1 from A - 120 sec		C read 1 from A - 28 sec (same %s)
C read 1 from B - 126 sec		C read 1 from B - 28 sec

C read 2 from B - 60 sec		C read 2 from B - 15 sec
C read 2 from A - 58 sec		C read 2 from A - 15 sec

[ server writes to server ]		[ server writes to server ]

A write 4 to B - 210 sec		A write 4 to B - 36 sec (>55% net util)

[ server reads from server ]		[ server reads from server ]

A reads 4 from B - 216 sec		A reads 4 from B - 200 sec (delayed writes)

-------------------			-------------------

Private 10MB shared ethernet subnet	Private 100MB shared ethernet subnet

[ client writes to server ]		[ client writes to server ]

D write 1 to A - 153 sec		D write 1 to A -  78 sec (~5% net util)
D write 1 to B - 163 sec		D write 1 to B - 405 sec (0% net util!)

D write 2 to B -  74 sec		D write 2 to B - 193 sec (!)
D write 2 to A -  73 sec		D write 2 to A -  36 sec

D write 3 to A -  48 sec		D write 3 to A -  30 sec
D write 3 to B -  49 sec		D write 3 to B -  132 sec (!)

[ client reads from server ]		[ client reads from server ]

D read 3 from B - 49 sec		D read 3 from B - 18 sec (~35% net util)
D read 3 from A - 47 sec		D read 3 from A - 17 sec (~35% net util)

D read 1 from A - 142 sec		D read 1 from A - 51 sec (same %s)
D read 1 from B - 148 sec		D read 1 from B - 52 sec

D read 2 from B - 71 sec		D read 2 from B - 26 sec
D read 2 from A - 68 sec		D read 2 from A - 25 sec

-------------------

Observations:

01 - From an NT client, there is very little noticeable difference between
reading and writing from/to a NT native or SAMBA implementation of SMB/CIFS.

02 - SAMBA utilized bandwidth better than NT server when an NT client
writes to it on 
a 100MB pipe.

03 - Raid 5 is the NT server's bottleneck when reading on a 100MB pipe.

04 - A windows 98 client acts brain dead when writing to a SAMBA server on
a 100MB 
pipe (much worse than writing to a SAMBA server on a 10MB pipe)!

05 - Neither disk nor memory caching seemed to come into play in these tests.

06 - This test did not study multi-user read/writes, possible memory
problems with 
SAMBA when it forks a smbd for each user?

07 - This test did not study network congestion, read/write retries, or
multiple 
client versions connecting at once.

-------------------

Results:

Samba 2.0.3 is a pretty slick implementation of SMB/CIFS.  The price can't
be beat 
(free), but Samba will require more administration/setup since it's
interface is much 
more difficult to use than the PnC (point and click) gui of NT.  While this
may be 
daunting to NT PnC admins, Samba also gives fine grain access to the
mechanisms of 
SMB/CIFS for fine tuning.  For the purposes of read only file services, MAC
OSX and 
SAMBA seem to be a fine match against NT server native SMB/CIFS.  The only
outlier is 
the bizarre result with the 98 clients on the 100MB pipe.  More extensive
testing 
need to be done at the level of Apple Computer to reproduce business
standardized 
results.  Apple should really push the easy of use of MAC OSX and the lower
price 
point when considering NT's proclivity to BSOD weekly, the requirement to
buy CALS, 
and its constant security problems.

-------------------

- Marc DeBonis (Marc.DeBonis at vt.edu)
written 990525 v1.0
Virginia Tech / AIS-TS

-- 
Marc 'Doc' DeBonis      [Email: Marc.DeBonis at vt.edu (PGP/SMIME on rqst)] 
Programmer Analyst, Sr. [AIS-Technical Support Virginia Tech (VPI&SU)]
"Absit prudentia nil rei publicae profitur"
"Alterius non sit qui suus esse potest"
-\|/-