BUGs in Samba/CIFS service config UI

[xbmcg]

Hi there,

I try to tweek my installation to get a better R/W Performance and stumbled about the folowing switch, that cannot be set in the UI (setting it is forgotten after save)

ENABLE SENDFILE


Another issue is the usage of the Switch

Asynchronous I/O (AIO).

When turned on, on writing a large file to the NAS the write prozess breaks down with connection lost error and cannot be retried successfully. Turning off AIO writes OK.

[raulfg3]

Another issue is the usage of the Switch

Asynchronous I/O (AIO).

When turned on, on writing a large file to the NAS the write prozess breaks down with connection lost error and cannot be retried successfully. Turning off AIO writes OK.

only if you use SMB2 if you use NT1 works fine, is a good know limitation and is posted on the readme:

Code: Select all

Known issue
===========
SMB2 of samba 3.6.x have a bug, if you have a trouble, please use NT1 instead.

CIFS/SMB, This is the only way for now to get no file transfers aborts.

Here you have the options to use for now!

1- Disable AIO when using SMB2
2- Enable AIO , but use NT1 instead of SMB2

[xbmcg]

Thank you. So it is not fixed yet. There are recommendations to turn AIO on for performance improvement.
Does it mean, that NT1 with AIO is faster / better than SMB2?

[raulfg3]

YES for me, you must test both options ( NT1 & AIO and SMB2 without AIO)

[xbmcg]

Thank you.

I have tested both, seems NT1 is better (less jitter on transfer and slightly faster with sendfile and AIO enabled), still too slow (300Mbps on a 1G uplink).
I have noticed, that anyway when writing a large file, the transfer drops to zero on every 1GB chunk (see picture below). When Using CIFS on Windows
as target I get 900Mbps and constant speed for the whole file. Is there a special (zfs-mirror) reason for the performance drops every 1GB?

[raulfg3]

play with RX and TX buffers of SAMBA ( I reduce it to solve your problems to 32K)

[kernow]

I have that issue with large transfers just giving up, also. Hmm.