Re: [Dataspace-sabul] AMD64 Question.

[Yunhong Gu <yu...@la...>]

Hi, Steven,

On Tue, 24 Feb 2004, Steven M. Carter wrote:
> That would be odd:
> 
> [scarter@unet1 scarter]$ ls -l
> total 36885892
> -rw-rw-r--  1 scarter scarter 10485760000 Feb 19 15:47 10g_foo
> -rw-rw-r--  1 scarter scarter  5242880000 Feb 22 22:14 5g_datafile
> -rw-rw-r--  1 scarter scarter  2147483647 Feb 24 18:03 5g_foo
> -rw-rw-r--  1 scarter scarter 13239889920 Feb 19 15:42 datafile
> 
> 5g_foo is the file that was aborted.  There is a 10g and 13g file in that 
> same directory.  The OS had no problem with those.
> 
> Tsunami transfered a 10g file with no problems :)

Sorry, I checked the "sendfile" code and found that I misused the type of 
"streamsize", which I thought can be used for 64-bit file size (and 
somehow I remember that I did use this program to transfer large files). 
Anyway, I corrected it and upload a new file to the CVS of sourceforge, 
and you can check out the most recent source code there. Or, you can 
simply modify the "sendfile.cpp" file by yourself:

in line 71, change 
streamsize size = ifs.tellg();
to
unsigned long long size = ifs.tellg();
or
streampos size = ifs.tellg();


> It now uses 9000 byte packets (I see 9014 on the wire as opposed to 1514),
> but the transfer rate goes from ~830Mb/s to ~395Mb/s.

I don't understand this either. However, my experience on using 9000byte 
MTU is very little. I read your testing report on both UDT and other 
protocols, including iperf, it seems 830Mbps is normal for the link you 
are testing, so I think you can continue to use 1500byte setting. 

From my experiences, 9000MTU only has significant effect on 10GigE links. 
On 1GigE links and below, 1500byte MTU works fine. On the other hand, even 
if you set 9000 UDT_MTU option on a 1500 MTU link, the performance should 
not degrade that much. It can be related to hardware/OS though. 

> 
> > For the file corruption error, it may be caused by a bug in UDT that we 
> > already fixed. You can find a new version of UDT release (v1.2) at
> > sourceforge.net/projects/dataspace.
> 
> I need to do a test from an IA32 machine to another IA32 machine to rule 
> that out.
> 

Please let me know if the file corruption error is still there.

Thanks,
Gu