On Wed, Oct 20, 2004 at 07:16:18PM +0100, Matt Clark wrote:
> >Hyperthreading is actually an excellent architectural feature that
> >can give significant performance gains when implemented well and used
> >for an appropriate workload under a decently HT aware OS.
> >IMO, typical RDBMS streams are not an obviously appropriate workload,
> >Intel didn't implement it particularly well and I don't think there
> >are any OSes that support it particularly well.
> >But don't write off using it in the future, when it's been improved
> >at both the OS and the silicon levels.
> You are quite right of course - unfortunately the current Intel
> implementation meets nearly none of these criteria!
Indeed. And when I said "no OSes support it particularly well" I meant
the x86 SMT implementation, rather than SMT in general.
As Rod pointed out, AIX seems to have decent support and Power has a
very nice implementation, and the same is probably true for at least
one other OS/architecture implementation.
> As Rod Taylor pointed out off-list, IBM's SMT implementation on the
> Power5 is vastly superior. Though he's also just told me that Sun
> is beating IBM on price/performance for his workload, so who knows
> how reliable a chap he is... ;-)
In response to
pgsql-performance by date
|Next:||From: Mark Wong||Date: 2004-10-20 22:05:28|
|Subject: Re: futex results with dbt-3|
|Previous:||From: Matt Clark||Date: 2004-10-20 18:16:18|
|Subject: Re: how much mem to give postgres?|