[Info-vax] Linux 40 GbE and 100 GbE NIC Performance

[Stephen Hoffman]

On 2015-01-25 01:36:55 +0000, David Froble said:

> I wasn't thinking "dedicated" from a HW perspective, rather more like 
> Galaxy, where the OS or something could automatically or be set to 
> assign cores to specific tasks.  Could be a lot of flexibility with 
> something like that.

FWIW, the dedicated lock manager has been available for many years.   
That was because the lock manager was — under heavy load, in some 
configurations — thrashing the system and particularly the caches.  
With a dedicated core, those problems are largely mitigated, up until 
the capacity of the dedicated core is exceeded.

For related details for applications that might wish to dedicate cores, 
see the OpenVMS class scheduler support documentation.

These are some of the issues that can arise with SMP scaling, and with 
cluster scaling.   Keeping the caches coherent adds overhead to the 
design, but folks using in an OpenVMS-style SMP box expect their caches 
to work; to remain coherent.  The Alpha architecture caused enough 
consternation with its comparatively loose write ordering expectations, 
for that matter.  The x86 write-ordering is rather more strict.

> Some people might be doing no network activity, and, some people might 
> be doing mostly network activity.

That doesn't begin to describe how far apart different applications 
are.  Some OpenVMS configurations with SMP and applications, and some 
combination of clustering and/or network activity and/or I/O loading 
will work fine at 16 and variously 32 cores, and some other 
applications will run into the proverbial wall at eight or twelve 
cores, or some even with even a second core.



-- 
Pure Personal Opinion | HoffmanLabs LLC