Re: Help!!

On Wed, 5 Jul 2000, Merton Campbell Crockett wrote:

> On Wed, 5 Jul 2000, David Luyer wrote:
>
> > > On Mon, 3 Jul 2000, jubaco30 wrote:
> > > > I need to know urgently(An Administrative Decision) if squid takes
> > > > advantage of SMP (Redhat 6.2 with 2 processors). The main questions is
> > > > if the granularity of locks in I/O lead to similar performance(with
> > > > smp, without smp)?.
> > >
> > > It is my understanding that Squid is implemented as a single process with
> > > its own internal state machine/executive to manage its activity.
> >
> > Under Linux it multi-threads if you enable ASYNCIO, which means it scales
> > much better over multiple separate disks (without RAID).
> >
> > > Under BSD/OS 4.x, Squid's apparent performce improved dramatically on
> > > multi-processor systems. The observed performance improvement on the dual-
> > > and quad-processor systems that we have is more a function of how well the
> > > OS manages its resources and activities in an SMP environment.
> >
> > Under BSD/OS you're likely to not be using ASYNCIO, so you have limited
> > gain from SMP for Squid, but there is still some gain there.
> >
> > > While we do have a few Linux systems, we tend to use BSD/OS where the
> > > leather meets the road. This is probably more a function of two decades of
> > > experience with the Berkeley Software Distribution from CSRG and BSDi.
> >
> > Well we use it too (as a result of policy, not of choice) but the number
> > of bugs and lack of development tools is just damn annoying to people
> > who are used to something as robust and well supported for development as
> > Linux. Fortunately Squid doesn't hit many of the bugs (but OTOH it doesn't
> > do ASYNCIO under BSD/OS which is a bit of a performance pain, but can be
> > countered by just buying faster boxes than you would need to with Linux).
>
> I gather from your comments that ASYNCIO is similar in nature to the BSD/OS
> file system's "soft update" feature. As I recall from a colleague's tests,
> the performance improvement from using "soft updates" eliminated the need
> to add a second processor to systems we were designing for our customers.

softupdates on freebsd increases the performance of create/write/destroy
operations by a factor of like 300 in my tests, that's the difference
between being able to 30 create/write/destroy cycles on the same file per
filseystem per second, and being able to do ~9000. I'm not real sure how
much smp would help in the situation, but as far as I'm concerned ufs
without softupdates is broken. A similar (but not identical) linux box I
have can do about 11,000 with async io turned on (the default) and while
I will say that ext2 is more brittle than ufs in the event of crash I'm
noot sure I care with proxy cache filesystems.

> Interestingly BSDi doesn't tout this feature all that much. I guess it
> doesn't have the sex-appeal of the SMP cachet.
>
> Merton Campbell Crockett
>

-- 
-------------------------------------------------------------------------- 
Joel Jaeggli				       joelja@darkwing.uoregon.edu    
Academic User Services			     consult@gladstone.uoregon.edu
     PGP Key Fingerprint: 1DE9 8FCA 51FB 4195 B42A 9C32 A30D 121E
--------------------------------------------------------------------------
It is clear that the arm of criticism cannot replace the criticism of
arms.  Karl Marx -- Introduction to the critique of Hegel's Philosophy of
the right, 1843.