Bad NCP Performance vs CIFS

Question

Our enviroment :
OES 11 SP1
Connect via 10Gbit Fibre (direct attached cable to switch)
Clients WINDOWS 7
Connect via 1 Gbit
We se Bad NCP perfomance on the same Volume vs CIFS
While copying a file from the workstation to a volume via NCP we get about ~45 MB/s
While copying a file from the workstation to the same volume via CIFS we get about ~100 MB/s
NCP can be slower but thats a huge difference, any ideas to step further are welcome.
regards

Answer 1

Originally Posted by mrosen
On 29.10.2013 21:16, Bob-O-Rama wrote:
>
> No idea.
>
> I know people have reported performance differences, in my experience
> its not been substantial. Perhaps we have some super awesome network
> or something. ;-) Which we do... because "network" is in my job
> title.
Does that mean you see ~ 100MB/sec via NCP on a GB conenction?
CU,
Massimo Rosen
Novell Knowledge Partner
No emails please!
Untitled Document
finally found the time to work a little on this. dedicated boxes (pretty old but decent hardware) on a dedicated LAN, no traffic apart from what i've triggered. the tests are far from real-life operations as it's just about copying a 3GB iso back and forth and doing something similar with "lan speed test lite" (totusoft.com). observations made as follows:
on XP the NCP operations outperform CIFS operations all across the board. i get constant 70MB/s on both reads and writes via NCP, CIFS writes at about 60 and reads 50MB/s (server and workstations rebooted after each test).
on W7 NCP performance is about the same as on XP, CIFS reads are slightly below 70MB/s but CIFS writes constantly at around 105MB/s.
there some pretty confusing things i wonder about:
in the past the FILE_COMMIT setting on the server just decided whether or not to honor client initiated commit requests (did this change?). copying a file with windows explorer never triggers a commit, so neither client nor server settings should be a factor here. in fact, the server setting really has no effect, but running W7 the client setting really makes a difference (though there's no NCP 59 verb leaving the client). the 70MB/s via NCP on W7 could only be reached with FILE_COMMIT being enabled on the client. while i set commit on and oplocks off all of my life anyway i'd be interested how setting commit off could drop transfer rates down to 60MB/s without a commit request on the wire regardless of this setting. obviously there are effects other than requesting a commit on a buffer-flush call made by an application. maybe something alan like to share with us...
what makes the W7 CIFS writes significantly faster? i don't think that caching or lazy writes are a factor here as tests with files larger than server's or workstations' memory lead to identical results, i.e. continuous data flow on the network and a max of 2 seconds I/O on the drives after the copy has finished. no difference on calling "sync" or "fsync" after NCP / CIFS operations. not sure how this integrates with NSS, though.
write caches on harddrives and array controller are disabled, btw.
finally: does anyone have an idea what the FILE_FLUSH parameter in cifs.conf is about? setting it yes or no didn't make a difference for these tests, just being curious.