[uml-devel] Fighting latency

[uml-devel] Fighting latency

[nils toedtmann]

Moin,

in some UMLs i run services which do not like latency, eg. VoIP- and gameserver. 
I previously observed my UMLs going into meditation for a split second every few 
seconds, but that bucking almost disappeard by taking the usual tuning steps 
(SKAS, sysemu, $TMP on tmpfs). Nevertheless, my UMLs are still not as responsive 
as i would like them to. It's not a matter of CPU or network performance. Host
and guest are >=95% idle all the time, bandwidth use is far from my UML-bench-
marks.

I tried to renice the UML processes, but seems not to help (it's hard to measure,
just the "feeling" while listening to audio or onlinegaming). 

Any suggestions?

I enabled "PREEMPT" in host kernel. Maybe that's a bad idea?

/nils.

-- 
there is no sig.


-------------------------------------------------------
This SF.Net email is sponsored by:
Sybase ASE Linux Express Edition - download now for FREE
LinuxWorld Reader's Choice Award Winner for best database on Linux.
http://ads.osdn.com/?ad_id=5588&alloc_id=12065&op=click
_______________________________________________
User-mode-linux-devel mailing list
User-mode-linux-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/user-mode-linux-devel

Re: [uml-devel] Fighting latency

[Blaisorblade]

On Thursday 04 November 2004 16:50, nils toedtmann wrote:
> Moin,
>
> in some UMLs i run services which do not like latency, eg. VoIP- and
> gameserver. I previously observed my UMLs going into meditation for a split
> second every few seconds, but that bucking almost disappeard by taking the
> usual tuning steps (SKAS, sysemu, $TMP on tmpfs). Nevertheless, my UMLs are
> still not as responsive as i would like them to. It's not a matter of CPU
> or network performance. Host and guest are >=95% idle all the time,
> bandwidth use is far from my UML-bench- marks.
>
> I tried to renice the UML processes, but seems not to help (it's hard to
> measure, just the "feeling" while listening to audio or onlinegaming).
>
> Any suggestions?
>
> I enabled "PREEMPT" in host kernel. Maybe that's a bad idea?

This problem is an old one. Search the archives for "odd ping problems", which 
is about long ping delay.

A fix was proposed a while ago, which consisted in mlock()ing the memory. But 
Jeff Dike refused it because he said "I want to write a much more complex and 
more optimal solution. Mlock() is an hack".

Even so, I'll try to update the patch, fix it on the security side (that one 
required setuid()'ing UML, which is purposeless) and push it for inclusion, 
if possible.

Bye
-- 
Paolo Giarrusso, aka Blaisorblade
Linux registered user n. 292729


-------------------------------------------------------
This SF.Net email is sponsored by:
Sybase ASE Linux Express Edition - download now for FREE
LinuxWorld Reader's Choice Award Winner for best database on Linux.
http://ads.osdn.com/?ad_id=5588&alloc_id=12065&op=click
_______________________________________________
User-mode-linux-devel mailing list
User-mode-linux-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/user-mode-linux-devel

Re: [uml-devel] Fighting latency

[Nuno Silva]

Blaisorblade wrote:

...

> A fix was proposed a while ago, which consisted in mlock()ing the memory. But 
> Jeff Dike refused it because he said "I want to write a much more complex and 
> more optimal solution. Mlock() is an hack".

Maybe having the linux binary in a ramfs will work better?
You can also do some creative mount --bind, to mount only one ramfs 
instance, if you have many UMLs running.

Regards,
Nuno Silva


-------------------------------------------------------
This SF.Net email is sponsored by:
Sybase ASE Linux Express Edition - download now for FREE
LinuxWorld Reader's Choice Award Winner for best database on Linux.
http://ads.osdn.com/?ad_id=5588&alloc_id=12065&op=click
_______________________________________________
User-mode-linux-devel mailing list
User-mode-linux-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/user-mode-linux-devel

Re: [uml-devel] Fighting latency

[Blaisorblade]

On Friday 05 November 2004 03:54, Nuno Silva wrote:
> Blaisorblade wrote:
>
> ...
>
> > A fix was proposed a while ago, which consisted in mlock()ing the memory.
> > But Jeff Dike refused it because he said "I want to write a much more
> > complex and more optimal solution. Mlock() is an hack".
Actually he said something much more complex than that. You can find his 
worries in this email.

> Maybe having the linux binary in a ramfs will work better?
Wise idea! But you forgot the UML "RAM" from $TMPDIR.

Ramfs data are not swappable, so using that both for the UML binary and for * 
$TMPDIR * should work.

However, Jeff wouldn't have a different opinion on this, but it can be done 
without patching UML.

In fact, he worries about the host going OutOfMemory or anyway actually 
needing hard to swap out memory; this could even happen when, for instance, 
UML is using all memory just for caching (and maybe caching unused data). In 
the meanwhile, while one UML caches an use-only-once file, the host goes 
OutOfMemory and kills one UML instance. He wanted the UMLs to release memory 
to the host when needed.

This, however, is not a problem for dedicated UML hosting if you don't 
oversell UML ram.

> You can also do some creative mount --bind, to mount only one ramfs
> instance, if you have many UMLs running.

Not needed I think - use different names for UML binaries as needed. It will 
pickup a good name for the RAM file.
-- 
Paolo Giarrusso, aka Blaisorblade
Linux registered user n. 292729


-------------------------------------------------------
This SF.Net email is sponsored by:
Sybase ASE Linux Express Edition - download now for FREE
LinuxWorld Reader's Choice Award Winner for best database on Linux.
http://ads.osdn.com/?ad_id=5588&alloc_id=12065&op=click
_______________________________________________
User-mode-linux-devel mailing list
User-mode-linux-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/user-mode-linux-devel

Re: [uml-devel] Fighting latency

[Nuno Silva]

Blaisorblade wrote:
>>Maybe having the linux binary in a ramfs will work better?
> 
> Wise idea! But you forgot the UML "RAM" from $TMPDIR.

No i didn't :-) And i advise against it because of the problems that you 
point out below (OOM). Anyway, if the data is frequently used it won't 
be sent to swap... If you only touch that data once every 5 minutes and 
you have some memory pressure you may get a bigger latency.

If you must have fast-as-it-gets memory access making several UBDs, one 
of them very small in ramfs, may be better, YMMV.

...

>>You can also do some creative mount --bind, to mount only one ramfs
>>instance, if you have many UMLs running.
> 
> Not needed I think - use different names for UML binaries as needed. It will 
> pickup a good name for the RAM file.

I was thinking about chroot environments. And saving some MB by using 
the same binary - living in ramfs - to run many UMLs.

Regarding double-cacheing you may try NFS or directIO (with directIO the 
guest can create disk I/O even if it's not needed and ruin the host's 
I/O performance). Again, YMMV.

Thanks,
Nuno Silva


-------------------------------------------------------
This SF.Net email is sponsored by:
Sybase ASE Linux Express Edition - download now for FREE
LinuxWorld Reader's Choice Award Winner for best database on Linux.
http://ads.osdn.com/?ad_id=5588&alloc_id=12065&op=click
_______________________________________________
User-mode-linux-devel mailing list
User-mode-linux-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/user-mode-linux-devel