[Info-vax] What choices are available to the OpenVMS Process Scheduler once all the CPUs are in use?

[lists at openmailbox.org]

On Thu, 11 Feb 2016 14:15:23 +0000
Kerry Main via Info-vax <info-vax at rbnsn.com> wrote:

> > -----Original Message-----
> > From: Info-vax [mailto:info-vax-bounces at info-vax.com] On Behalf Of
> > Carl Friedberg via Info-vax
> > Sent: 10-Feb-16 2:28 PM
> > To: info-vax at info-vax.com
> > Cc: Carl Friedberg <frida.fried at gmail.com>
> > Subject: Re: [New Info-vax] What choices are available to the OpenVMS
> > Process Scheduler once all the CPUs are in use?
> > 
> > Of course I know the 782 was ASMP, that was a (bad) joke. Hyper-
> > threading
> > is a nice way to up the marketing value by doubling the number of core
> > CPUs, but I doubt that VMS users will see much, if any benefit; and
> > possibly there will be a performance penalty. It can't hurt to test a
> > specific environment, as it would be hard to predict IMO.
> > 
> > Whatever the benefit of hyper-threading, I doubt anyone will ever
> > double
> > their throughput by turning it on. YMMV.
> > 
> > Carl
> 
> [snip..]
> 
> It's not just OpenVMS that sees little benefit with threads, the same
> is (mostly) true of other platforms. As I recall, the standard feedback
> from MS on the latest Windows Server versions is leave threads off.
> 
> When you mix threads with NUMA (most server HW today is NUMA 
> based - incl x86-64) and local vs. remote memory, to me, there is just
> way to many ways for things to go wrong performance wise.
> 
> To be fair - I did do a DC project last year where they were heavy users 
> of Solaris on the relatively current M5 SPARC servers. Their app was a 
> huge web site with lots of parallel independent web requests and they 
> used threads extensively. The M5 allowed up to 8 threads per core, but 
> the HW and OS were both designed to support these concurrent threads.

Indeed, a lot of the low-end T servers (T1000/T2000) are optimized for
thread throughput rather than high clocks per core. They're stinkers on a
lot of workloads (desktop, for example) but work well serving websites and
doing other work where having a lot going on overall is more important than
having any one thing going on fast.

> (case of single vendor making HW + OS ..remember the good old days?)

That's what I've been saying! :-)