[FUGSPBR] SMP e Polling (Era: OpenBSD/NetBSD/DragonFly)

[Giovanni P. Tirloni]

Patrick Tracanelli wrote:
> João Carlos Mendes Luís wrote:
> 
>>     Está rolando uma discussão sobre isso na -hackers, onde se discute 
>> que isso é teórico, e que tem gente usando DEVICE_POLLING com SMP sem 
>> problemas.
> 
> 
> Além do config "gritar" quando você tem ambos no kernel, quando eu 
> forcei mante-los o sistema... adivinha? hehehe :P

  Eu não acompanhei a -hackers esses dias mas na -performance falaram 
disso também (talvez com cópia para lá).

  Também já vi relatos de gente usando o polling com SMP sem 
problemas. Agora não sei era antes de implementarem o KSE ou depois.

   A mensagem abaixo parece bastante esclarecedora (se o autor estiver 
correto). Vamos ver se o Luigi responde ela na lista.

-- 
Giovanni P. Tirloni


From: Jason Stone <freebsd-performance em dfmm.org>
To: freebsd-performance em freebsd.org
Cc: luigi em iet.unipi.it
Message-ID: <20041021133719.S79820 em walter>
Subject: Re: decreasing interrupt CPU load

 > If you look at http://info.iet.unipi.it/~luigi/polling, the last Q &
 > A question suggests why it is disabled for SMP.  It seems that
 > polling only runs on one thread whereas an smp box might handle
 > concurrently interupts from different devices.
 >
 > Can the scheduler move the thread to another cpu or is it locked on
 > a particular cpu?

thanks for the pointer.  it seems to me that the thread doing the 
polling could move from cpu to cpu, but that's not the issue - the 
issue, if I'm understanding the author, is that the polling thread 
will always be a single thread, whereas if we use traditional 
interrupts, they can be handled on multiple cpu's concurrently.  so 
interrupts, with multiple cpu's to handle them, might give much better 
performance than a single polling thread.

if that's the only issue, then preventing one from compiling with both 
SMP and DEVICE_POLLING doesn't seem necesary, since you can turn 
polling on and off at runtime with sysctl's.

luigi: does that sound right to you?  has any consideration been 
giving to removing the restriction preventing DEVICE_POLLING from 
building on an SMP system, in light of the fact that polling can just 
be turned off if it's giving suboptimal performance?

  -Jason

_______________________________________________________________
Para enviar um novo email para a lista: fugspbr em fugspbr.org
Sair da Lista: http://lists.fugspbr.org/listinfo.cgi
Historico: http://www4.fugspbr.org/lista/html/FUG-BR/