[AG-TECH] Mac buffer tweak?

[gurcharan khanna]

All,

To be precise, the only reason I have manipulated these parameters is to 
be able to render
the 19.2 Mbps mpeg2 stream from Seattle (233.0.73.26) with VLC and 
similar higher throughput
streams. I never had a problem with vic streams, so my experiments and 
observations
reflected the need to optimize at higher bitrates.

However, I also played around with these settings at home with my 
wireless connection via Airport
to roadrunner cable modem internet access, so I was interested in 
optimizing both tcp and
udp. I also seem to recall something about the maxsockbuf being large 
enough to
accommodate the tcp send/recv space settings, but perhaps that was 
inaccurate.

Fyi,

-gurcharan

Piers O'Hanlon wrote:
> Hi,
>
> I think the key sysctl here is this one: kern.ipc.maxsockbuf=8000000
> which determines the max UDP buffer sizes (the others, as Chris points
> out, are only relevant to TCP).
>
> What would be useful to check is to run netstat -s on your system with
> and without the ker.ipc.maxsocketbuf setting - and look at the UDP
> section on "dropped due to full socket buffers" and see how this
> changes.
>
> However this is odd as vic does set the socket buffer sizes in the
> application. One explanation maybe that the application provided
> buffer sizes are a guide and if they're overrun then it can use more
> memory but if the default max (on my MacBook they're set to 262144) is
> too low then they will run out and hence you'll get bad video.
>
> The other thing that is strange is that this doesn't appear to happen
> on Linux (maybe it is and we haven't noticed it) but after a quick
> look at the Linux kernel it seems that the SO_RCVBUF setsocketopt()
> kernel implementation doubles the supplied value which may makes
> things ok on Linux? However the max buffer size in Linux (Ubuntu 7.10
> anyway) is /proc/sys/net/core/wmem_max 131071 (which is lower than
> OSX)...
>
> If we find that vic's default's are too low then we can tweak them,
> though it may only be fixed on some systems but increasing the system
> maximum.....
>
> It maybe worth keeping an eye on UDP buffer overruns (use netstat -s)
> to see if this a more general problem.
>
> There's a good guide on UDP buffer sizing here:
> http://www.29west.com/docs/THPM/udp-buffer-sizing.html
>
> Piers.
>
>
> On 10/11/2007, gurcharan khanna <gskpop at cis.rit.edu> wrote:
>   
>> Christoph Willing wrote:
>>     
>>> On 10/11/2007, at 7:56 AM, gurcharan khanna wrote:
>>>
>>>       
>>>> Mark,
>>>>
>>>> my link to the source for this seems to be broken but here's what i
>>>> have personally on my mac.
>>>>
>>>> in /etc/sysctl.conf, add these lines:
>>>>
>>>>
>>>> net.inet.tcp.recvspace=4000000
>>>> net.inet.tcp.sendspace=4000000
>>>> kern.ipc.maxsockbuf=8000000
>>>>         
>>> Those settings look like they may have come from:
>>>     http://www.psc.edu/networking/projects/tcptune/
>>>       
>> You're right about the URL. However, the link doesn't work anymore....
>>     
>>> However, note that TCP tuning is the subject there, not UDP (which is
>>> what multicast uses).
>>>       
>> Understood. However, changing these settings DOES in fact improve things
>> considerably, predictably,
>> and repeatedly. I don't quite understand it, but it does work.
>>
>> -gurcharan
>>     
>>> chris
>>>
>>>
>>>
>>>       
>>>> you probably have to restart to activate them. you can set these from
>>>> the command line as well but they won't "stick", but you can try out
>>>> different
>>>> settings that way.
>>>>
>>>> use: sysctl -w net.inet.tcp.recvspace=4000000        (for example)
>>>>
>>>> to set these values interactively.
>>>>
>>>> -gurcharan
>>>>
>>>>
>>>>
>>>> Leonard, Mark wrote:
>>>>         
>>>>> Listers--
>>>>>
>>>>> It was probably a year ago that a setting change for Mac OS that
>>>>> improved video quality of big video multicast streams (Gurcharan?)
>>>>> but I
>>>>> can't seem to find it any where...
>>>>>
>>>>> Can someone repost or point me towards a troubleshooting page for
>>>>> tweaks
>>>>> like this?
>>>>>
>>>>> Thanks
>>>>> Mark
>>>>>
>>>>>
>>>>> Mark Leonard
>>>>> DL Studios, UNH
>>>>> 603-862-0192
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>           
>>>> --
>>>> -------------------------
>>>> Gurcharan S. Khanna, Ph.D.
>>>> Director of Research Computing
>>>> Office of the Vice President for Research
>>>>
>>>> Director, Interactive Collaboration Environments Laboratory,
>>>> Center for the Advancing the Study of Cyberinfrastructure
>>>> ---
>>>> Rochester Institute of Technology
>>>> 1 Lomb Memorial Drive
>>>> Rochester, New York 14623-5603
>>>> Phone: 585-475-7504  ~  Cell: 585-451-8370
>>>> Email: gurcharan.khanna at rit.edu
>>>> Http:  people.rit.edu/gskpop
>>>>
>>>>         
>>> Christoph Willing                       +61 7 3365 8350
>>> QCIF Access Grid Manager
>>> University of Queensland
>>>
>>>
>>>
>>>       
>> --
>> -------------------------
>> Gurcharan S. Khanna, Ph.D.
>> Director of Research Computing
>> Office of the Vice President for Research
>>
>> Director, Interactive Collaboration Environments Laboratory,
>> Center for the Advancing the Study of Cyberinfrastructure
>> ---
>> Rochester Institute of Technology
>> 1 Lomb Memorial Drive
>> Rochester, New York 14623-5603
>> Phone: 585-475-7504  ~  Cell: 585-451-8370
>> Email: gurcharan.khanna at rit.edu
>> Http:  www.rit.edu/~gskpop
>>
>>
>>     


-- 
-------------------------
Gurcharan S. Khanna, Ph.D.
Director of Research Computing
Office of the Vice President for Research

Director, Interactive Collaboration Environments Laboratory,
Center for the Advancing the Study of Cyberinfrastructure
---
Rochester Institute of Technology
1 Lomb Memorial Drive
Rochester, New York 14623-5603
Phone: 585-475-7504  ~  Cell: 585-451-8370
Email: gurcharan.khanna at rit.edu
Http:  www.rit.edu/~gskpop