NAT: CIFS connection to host is terribly slow

[Artem S. Tashkinov]

Probably NAT is not meant for transferring data between host OS and guest OS, but since it's possible I suppose such a connection should have a speed comparable to Host-only connection.

However it's not the case.

When copying data from Fedora 14 guest to Fedora 14 host (both have i686 arch) I get 196KB/sec connection speed on average. Both hosts are 100% idle when copying (host OS exports SMB shares, guest OS has them mounted via CIFS).

When I copy data using Host-only adapter the speed soars to 20-30MB/sec.

P.S. Guest additions are not installed. Guest OS has Intel PRO/1000MT Desktop (82540EM) adapter.

[Artem S. Tashkinov]

Low speed when copying a file from a guest OS to a host OS - screenshot

[vasily Levchenko]

Could you please attach the log file?

[vasily Levchenko]

could you please attach output of ​ffsb for NAT and host-only?

[Artem S. Tashkinov]

Please, give me a ffsb profile to test. There are eight exemplary ones, I've no idea which one to use.

[Artem S. Tashkinov]

NAT'ted VBox.log

[Artem S. Tashkinov]

Host-only VBox.log

[vasily Levchenko]

Replying to birdie:

Please, give me a ffsb profile to test. There are eight exemplary ones, I've no idea which one to use.

please use profile_everything config with replacing location property with path to some test folder on smb/cifs share.

[filesystem]
   location = 

[Artem S. Tashkinov]

ffsb NAT results

[Artem S. Tashkinov]

ffsb host-only results

[Artem S. Tashkinov]

I don't know how these results can be helpful - since that's exactly what I've said in the first post.

[vasily Levchenko]

Replying to birdie:

I don't know how these results can be helpful - since that's exactly what I've said in the first post.

Benchmark is more obvious to simplify the test case than real application (in case if benchmark has the same behavior as application ofc). Particular this benchmark do several operations, and more flexible with initial parameters, and shows wider specter of CIFS related issues.

BTW: If the behavior of mc is the same for NFS share for you?

[Artem S. Tashkinov]

I did a better and cleaner test (while using NAT connection):

On host:

$ nc -l 1234 > file.temp

On guest

$ time nc 10.0.2.2 1234 < /proc/kcore real : 1m10s

kcore weighs 1,032,188K, so we're getting 14,745KB/sec speed which is (I suppose) close to the theoretical maximum since guest is eating 100% of the host CPU at this moment.

So, actually I don't know what to do with this bug report - it's 100% reproducible using CIFS connection, it doesn't hold true when I use direct TCP connection (e.g. using nc/netcat).

[vasily Levchenko]

Replying to birdie:

So, actually I don't know what to do with this bug report - it's 100% reproducible using CIFS connection, it doesn't hold true when I use direct TCP connection (e.g. using nc/netcat).

I'll just change the header to mark as CIFS specific.

[vasily Levchenko]

Replying to birdie:

So, actually I don't know what to do with this bug report - it's 100% reproducible using CIFS connection, it doesn't hold true when I use direct TCP connection (e.g. using nc/netcat).

Does it look host/guest specific, e.g. you don't see the same effects with CIFS on other guests or hosts? Here on Mac and Ubuntu hosts I don't see such dramatical performance slowdown on CIFS.