Re: hyperthreading

joel-garry_at_home.com (Joel Garry) wrote in message news:<91884734.0401121158.42cf7609_at_posting.google.com>...
> Mladen Gogala <mgogala_at_adelphia.net> wrote in message news:<pan.2004.01.11.06.15.24.204294_at_adelphia.net>...
> > On Fri, 09 Jan 2004 14:42:18 -0800, Joel Garry wrote:
> >
> > >
> > > Now all we need is the 2.6 kernel and Oracle properly threaded to use
> > > it.
> >
> > You would like to replace fork-ed processes with threads?
> > Why would that be good? The problem is that threads all share
> > the same address space while processes have their own address
> > spaces. Shared are only certain critical sections which are
> > known as "Shared Global Area". I should also add that the threads
> > implementations are OS specific (ranging from light-weight processes
> > to user mode thread which are actually serially executed in such a
> > way that user thinks that threads are run concurrently. The latter are
> > know as "green threads".). Using threads instead of fork & SYSV IPC would
> > probably have disastrous consequences for performance. It would turn even
> > the best Unix systems into an Xtreme Perversion, XP for short.
>
> I was hoping there would be coding that could be done on the Oracle
> side to take advantage of hyperthreading. The idea is to use hardware
> capabilities to improve things (I would speculate that such things as
> SGA buffer maintenance could benefit from this). If you get 30%
> improvement by using these capabilities, after the expense of a 50%
> decrease simply moving away from forks, that would indeed be silly.
> But if we're going to compete with other products optimized to use
> perversion, we must look at how everything works together and not be
> religious about ... whatever.
>
> jg

there was a recent posting on 2cpu.com regarding benchmarking dual Xeon 2.8 GHz MP cpus vs single 3.2 GHz P IV.

http://www.2cpu.com/articles/ht_explored/index.html

The most impressive point had to do with memory bandwidth, as the dual Xeon MP at 2.8 GHz was stomped by the single 3.2 GHz chip due to the faster front side bus (400 MHz vs 533 MHz). Also noteworthy was the author's point that its likely that hyperthreading was thrashing the level 1 and level 2 caches, in some cases causing loss of performance relative to a non-hyperthreaded configuration.

If the case that you are attempting to benchmark is use of the parallel query option where you are performing full table scans and frequently waiting on physical IO to take place, hyperthreading sounds like it would help, as while one thread is stalled the processor is busy executing on the other logical CPU.

If you have a seriously heinous nested query running on both the physical and logical CPUs that performs no physical IOs and is completely memory bound (BCHR setter), I would think would provide a little or negative performance increase.

guess the only way to find out is to benchmark them. I do have access to a Dell PE 2650 3.06 GHz 1 MB L3 cache server this week that has 9.2.0.4 loaded on both w2k3 server and RH 8.0, but that's not happening for me this week. Maybe next weekend.

Pd Received on Mon Jan 12 2004 - 19:54:54 CST