CIFS read-ahead

[Nicholas M. Kirsch]

> 1. I can get 100-114MB/s from Samba 3.0.0beta1 or whatever when I use
> cifs_bm with a stream of read&x requests. [1]

This is great -- it gives me hope -- the question is whether a stock
Windows client can be tuned/massaged to perform so well.

> 2. Windows never seems to achieve those rates because it inserts all sorts
> of additional commands between reads.

Hmm. Sounds less promising.

> 1. Are the files cached?

The files are being served from a 2GB ramdisk in order to eliminate the
disk overhead.

> 2. Were you simply using the Windows exploder to copy the files and do you
> have a trace?

I was not using Windows exploder, but a very simple application that just
has read() calls in a loop -- essentially the same as exploder, I imagine.
I do have a trace -- it is 2 MB in size, would you like me to send it to
you personally?

> 3. What happens if you use a series of ReadFile calls in a simple C
> program?

I have not tried ReadFile, but ReadFileEx and async IO with no noticable
improvement in results. Your message indicating a syncronous redirector
seems to be the culprit.

> BTW, there is an industry standard. Mb/s is megabits/sec, while MB/s is
> megabytes/sec.

You are correct, thank you.

Nick