Re: SMP performance on drgonfly

[Kris Kennaway <kris@xxxxxxxxxxxxxx>]

On 2007-05-19, Erik Wikström <erik-wikstrom@telia.com> wrote:
> On 2007-05-19 01:58, Kris Kennaway wrote:
>> On Fri, May 18, 2007 at 04:46:27PM -0700, Matthew Dillon wrote:
>>>     A large chunk of the kernel still runs under the big giant
>>>     lock, including the light weight processes that libthread_xu
>>>     uses, so something like mysql is going to hit a lot of BGL
>>>     contention.
>> 
>> Oh, OK.
>> 
>> What subsystems are out from under the big giant lock, so I can look
>> for another benchmark to compare with?  Also, what profiling and
>> contention measurement tools do you have, so I can try to confirm that
>> this is the issue?
>> 
>>>     You may be able to get DragonFly to run on the machines you
>>>     were having problems with by compiling it with SMP but 
>>>     without APIC_IO.  With that combination DragonFly will use
>>>     the PIC in SMP mode, which usually works.
>> 
>> The issue is that I cannot even install with a UP kernel, so I can't
>> recompile to test this.  I posted with more details about this problem
>> a few months ago.
>
> You should be able to build a new install CD, with the needed kernel 
> options, on the computer where you got it running. Check out the 
> nrelease framework for more details.

I dont think it is likely that Matt's suggestion will work given that
UP kernels do not boot, indicating it's not an SMP APIC issue.  Also
now that I have this other machine running I don't currently have a
motivation to try on 8-core AMD machines.

I was hoping that Matt would reply to my questions above so I could
try to find a more level playing field on which to compare Dragonfly
and FreeBSD, but I'm afraid that the answer is that there currently
are *no* kernel subsystems that are not Giant-locked, so all workloads
will perform poorly on DragonFly (and there are no specialized tools
for studying SMP behaviour -- certainly I couldn't find any when I
looked).

Can someone confirm whether this is true?

Kris