RFC: How to fix the NFS/iSCSI vs TSO problem

[araujobsdport at gmail.com]

Hello Rick,

I have some production servers with lots of NFS users that I can make a test too, but it will cost time and we can only verify regression, as I can't make any benchmark there!

Let me check, how loaded is this server and I tell you later; if you want me do more tests, I can do so.

Best Regards.

On 2014/4/1, at 8:41, Rick Macklem <rmacklem at uoguelph.ca> wrote:

> Jordan Hubbard wrote:
>> 
>> On Mar 31, 2014, at 8:53 AM, Marcelo Araujo <araujobsdport at gmail.com>
>> wrote:
>> 
>>> I understand your concern about add more one sysctl, however maybe
>>> we can
>>> do something like ZFS does, if it detect the system is AMD and have
>>> more
>>> than X of RAM it enables some options by default, or a kind of
>>> warning can
>>> be displayed show the new sysctl option.
>>> 
>>> Of, course other people opinion will be very welcome.
>> 
>> Why not simply enable (conditionally compile) it in only for the x64
>> architecture?   If you’re on a 64 bit Intel architecture machine,
>> chances are pretty good you’re also running hardware of reasonable
>> recent vintage and aren’t significantly HW constrained.
> I'm actually typing this on a single core amd64 with 2Gbytes of RAM, so
> I think enabling it only for both 64bits and at least some # of Gbytes of
> RAM would be better. (I agree that most amd64s will be relatively big
> machines, but not all;-)
> 
> My biggest problem is that I have no way of testing this on a fairly
> big amd64 server at this time and I'd be a lot more comfortable committing
> a patch that has been tested this way. (I realize that Marcelo has been
> running it for his benchmarks and that's a good start, but it isn't the
> same as a heavily loaded server.)
> 
> I notice that Alexander is on the cc list and I've added Garrett, since
> those are the two guys that have been doing a bunch of server testing
> (and my thanks go to them for this). Maybe they will have a chance to
> test this patch on a heavily loaded server?
> 
> Since I do want to test/debug the if_hw_tsomaxseg patch I have, I plan
> on inquiring to see if I can use something like the netperf cluster
> for this testing (in a couple of weeks when I get home).
> 
> rick
> 
>> I think it’s also fair to say that if you’re providing NFS or iSCSI
>> services on an i386 with 512M of memory or a similarly endowed ARM
>> or PPC system, performance is not your first and primary concern.
>> You’re simply happy that it works at all. ;-)
>> 
>> - Jordan
>> 
>>