mysql scaling questions

[Kris Kennaway]

Vlad GALU wrote:
> On 1/1/08, Kris Kennaway <kris at freebsd.org> wrote:
>> Gergely CZUCZY wrote:
>>> On Tue, Jan 01, 2008 at 05:04:56AM +0100, Kris Kennaway wrote:
>>>> Ivan Voras wrote:
>>>>> Kris Kennaway wrote:
>>>>>> Gergely CZUCZY wrote:
>>>>>>>> It looks like myisam is doing huge numbers of concurrent reads of the
>>>>>>>> same file which is running into exclusive locking in the kernel
>>>>>>>> (vnode interlock and lockbuilder mtxpool).  Does it not do any
>>>>>>>> caching of the data in userspace but relies on querying into the
>>>>>>>> kernel every time? innodb doesn't have this behaviour.
>>>>>>> Sorry, but was this a rethorical kind of question, or was this
>>>>>>> addressed to me? :)
>>>>>>> If the later, then how do I find this out?
>>>>>> It's a general question.  It looks like myisam either has a design
>>>>>> deficiency in this regard or it has poor defaults.  If it can be made to
>>>>>> improve caching of the data in userland then performance should improve.
>>>>> Isn't this common for software developed for Linux? I mean assuming
>>>>> syscalls are cheap; for example: gettimeofday(2), settitle(2), etc. I
>>>>> don't think the applications should be blamed for relying on performance
>>>>> optimizations not present in FreeBSD. Saying applications must do their
>>>>> own caching instead of relying on the kernel and need to avoid
>>>>> concurrent accesses to the same file seems like a doctrine from the dark
>>>>> ages.
>>>> Why?  Even if Linux magically has faster syscalls somehow, they are still not zero cost so avoiding huge numbers of unnecessary trips
>>>> into the kernel is in no sense a "doctrine from the dark ages".  Besides, if my hypothesis about the problem is correct then mysql
>>>> itself does this with the alternate innodb backend anyway.
>>> There's this SYSCALL CPU extension with the SYSENTER/SYSEXIT features. IIRC
>>> Linux takes advantage of this, while FreeBSD doesn't. I might be wrong here,
>>> of course.
>> FreeBSD does on amd64.  It still doesn't make syscalls free, so the
>> architectural principle of "cache data close to where it is needed"
>> continues to apply.
>>
>> Anyway, it remains to be understood whether linux really does have
>> faster syscalls, i.e. to understand exactly what unixbench is reporting
>> when it emits pretty numbers.  For example, how is it determining
>> "syscall overhead"?  Often this is done by calling a syscall that the
>> microbenchmark assumes is doing almost no work in the kernel.  This is
>> often chosen to be getpid() which may well be NULL on Linux, but
>> actually does do work on FreeBSD unless you remove COMPAT_43BSD from
>> your kernel.  Also I believe glibc caches getpid() in libc (again that
>> pesky architectural principle) so you need to be careful you are
>> actually doing the syscalls you think you are.
>>
> 
>    BTW, now with COMPAT_43 gone out of GENERIC, is it necesary to keep
> COMPAT_43TTY, even when Linux emulation is not needed?

I think a lot of old software (e.g. in ports) still uses it although 
someone is working on converting them to less archaic APIs.

Kris