The CIFSbench package

[David Collier-Brown]

Dan Kaminsky wrote:
> Jeremy, feel free to correct me on this but the ideal benchmark isn't merely
> CIFS but actual true Windows performance.  

	This was debated in the NFS world, due to concerns about
	client performance differences affecting what were supposed
	to be server benchmarks.  LADDIS ended up containing its
	own nfs code as a result.

	On the other hand, one often wants an end-to-end benchmark,
	as you suggest, so that the results from the benchmark are
	useful for sizing real systems.

	On the grippinmg hand, you want end to end, but you also
	want to measure what's server and what's client.


> 3)  Messages should be routed in the same way the user will end up using
> them, so *most probably* the clients should mount the SMB server as a drive.
> I don't know if this has a speed impact, though.

> Number three has a powerful impact--this means that we can fairly use
> something as simple as a disk benchmarker and use it to test Samba!  

	I'd suggest you're right, to get an end-to-end test,
	but that you also need the same load applied by a
	program making direct SMB calls, so that you can say how
	much the client (eg, NT) contributes.

	Not being a manufacturer, I don't really want compeditive
	performance numbers for servers.  I want to know how much
	load, measured in office-application users, a given server
	can handle, so I know when to upgrade. 

	I need to know how much the client slows us down, so I can
	know what "really fast" is.  Then I need to know how much
	the server can deliver, to see how close we can get to it
	for a given amount of money.
	 
--dave
-- 
David Collier-Brown,  | Always do right. This will gratify some people
185 Ellerslie Ave.,   | and astonish the rest.        -- Mark Twain
Willowdale, Ontario   | http://java.science.yorku.ca/~davecb
Work: (905) 477-0437 Home: (416) 223-8968 Email: davecb at canada.sun.com